Datasets:
Add files using upload-large-folder tool
Browse files- parse/train/86NHK__yFDl/86NHK__yFDl.md +322 -0
- parse/train/86NHK__yFDl/86NHK__yFDl_middle.json +0 -0
- parse/train/HJcSzz-CZ/HJcSzz-CZ.md +319 -0
- parse/train/HJcSzz-CZ/HJcSzz-CZ_content_list.json +1701 -0
- parse/train/HJcSzz-CZ/HJcSzz-CZ_middle.json +0 -0
- parse/train/HJcSzz-CZ/HJcSzz-CZ_model.json +0 -0
- parse/train/S1xSSTNKDB/S1xSSTNKDB_content_list.json +0 -0
- parse/train/S1xSSTNKDB/S1xSSTNKDB_model.json +0 -0
- parse/train/Sy3fJXbA-/Sy3fJXbA-.md +282 -0
- parse/train/Sy3fJXbA-/Sy3fJXbA-_content_list.json +1470 -0
- parse/train/Sy3fJXbA-/Sy3fJXbA-_middle.json +0 -0
- parse/train/Sy3fJXbA-/Sy3fJXbA-_model.json +0 -0
- parse/train/n4IMHNb8_f/n4IMHNb8_f.md +374 -0
- parse/train/n4IMHNb8_f/n4IMHNb8_f_content_list.json +2032 -0
- parse/train/n4IMHNb8_f/n4IMHNb8_f_middle.json +0 -0
- parse/train/n4IMHNb8_f/n4IMHNb8_f_model.json +0 -0
- parse/train/yILzFBjR0Y/yILzFBjR0Y.md +309 -0
- parse/train/yILzFBjR0Y/yILzFBjR0Y_content_list.json +1534 -0
- parse/train/yILzFBjR0Y/yILzFBjR0Y_middle.json +0 -0
- parse/train/yILzFBjR0Y/yILzFBjR0Y_model.json +0 -0
parse/train/86NHK__yFDl/86NHK__yFDl.md
ADDED
|
@@ -0,0 +1,322 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# TTT++: When Does Self-Supervised Test-Time Training Fail or Thrive?
|
| 2 |
+
|
| 3 |
+
Yuejiang Liu Parth Kothari Bastien van Delft
|
| 4 |
+
|
| 5 |
+
# Baptiste Bellot-Gurlet Taylor Mordan Alexandre Alahi
|
| 6 |
+
|
| 7 |
+
École Polytechnique Fédérale de Lausanne (EPFL)
|
| 8 |
+
|
| 9 |
+
{firstname.lastname}@epfl.ch
|
| 10 |
+
|
| 11 |
+
# Abstract
|
| 12 |
+
|
| 13 |
+
Test-time training (TTT) through self-supervised learning (SSL) is an emerging paradigm to tackle distributional shifts. Despite encouraging results, it remains unclear when this approach thrives or fails. In this work, we first provide an indepth look at its limitations and show that TTT can possibly deteriorate, instead of improving, the test-time performance in the presence of severe distribution shifts. To address this issue, we introduce a test-time feature alignment strategy utilizing offline feature summarization and online moment matching, which regularizes adaptation without revisiting training data. We further scale this strategy in the online setting through batch-queue decoupling to enable robust moment estimates even with limited batch size. Given aligned feature distributions, we then shed light on the strong potential of TTT by theoretically analyzing its performance post adaptation. This analysis motivates our use of more informative self-supervision in the form of contrastive learning for visual recognition problems. We empirically demonstrate that our modified version of test-time training, termed $T T T + +$ , outperforms state-of-the-art methods by significant margins on several benchmarks. Our result indicates that storing and exploiting extra information, in addition to model parameters, can be a promising direction towards robust test-time adaptation. Our code is available at https://github.com/vita-epfl/ttt-plus-plus.
|
| 14 |
+
|
| 15 |
+
# 1 Introduction
|
| 16 |
+
|
| 17 |
+
Machine learning models often struggle to generalize under distribution shifts. Even a perceptually mild shift between training and test data, e.g., JPEG compression, may cause severe prediction errors [1]. One popular family of methods to address this challenge is to learn an invariant representation across domains by making use of labelled training data and unlabelled test data simultaneously [2–5]. However, revisiting training data at test time can be impractical due to increasing privacy concerns, inflating sizes of datasets as well as many other real-world constraints. This shortcoming prompts a more challenging yet appealing test-time adaptation paradigm: given a trained model, how can we adapt it from one domain to another on the fly, without access to training data and human annotations?
|
| 18 |
+
|
| 19 |
+
One promising approach towards this goal is test-time training (TTT) through self-supervision [6]. The key idea of TTT is simple and straightforward: train the model on two tasks, namely a main task and a self-supervised learning (SSL) task, and update the model based only on the SSL task at test time. This technique implemented with self-supervised rotation prediction has shown encouraging results for improving the robustness of image classifiers under a variety of distributional shifts. Yet, its empirical performance is still inferior to other families of test-time algorithms [7, 8].
|
| 20 |
+
|
| 21 |
+
In this paper, we first take an in-depth look at TTT with emphasis on its limitations. Our analysis starts with a basic question: can TTT always mitigate the effects of distributional shifts? Through an illustrative problem, we show that the TTT framework can lead to surprising failures, deteriorating the test accuracy rather than improving it. This problem is largely attributed to the unconstrained update from the SSL task that interfere with the main task. To address this issue, we introduce a test-time feature alignment strategy by means of offline feature summarization and online moment matching: once training completes, we compute the mean and covariance matrix of training features and store them as part of the model, referred to as offline feature summarization; at test time, we encourage the test feature distribution to be close to the training one by matching the moments estimated online with those pre-computed offline, a process referred to as online moment matching.
|
| 22 |
+
|
| 23 |
+
One practical challenge for online feature alignment lies in scaling the strategy to problems with a large number of classes, as obtaining a robust estimate of moments often requires at least a handful of samples per class. To mitigate this issue, we draw inspiration from recent literature [9] and decouple the sample size from the batch size for moment estimates. Specifically, we maintain a large dynamic queue of encoded features and progressively update it in a mini-batch manner. This modification enables effective feature alignment even with limited batch size, greatly improving its viability in the online test-time setting.
|
| 24 |
+
|
| 25 |
+
Finally, we shed light on the strong potential of TTT through a theoretical analysis of the test accuracy after adaptation. In particular, we derive a lower bound of the test accuracy on the main task and show that it is expected to grow rapidly when the SSL task gets closer to the main task. These findings motivate our integration of contrastive representation learning [9–12], as a strong instance of SSL, into the TTT framework in visual recognition problems.
|
| 26 |
+
|
| 27 |
+
By combining the three proposed components, we devise an improved version of test-time training, termed $T T T + +$ . Experimental results show that $\mathrm { T T T } { + } { + }$ significantly outperforms other recent methods by significant margins on various robustness benchmarks. Our results suggest that exploiting extra information, including both task-specific information in the form of strong self-supervision and domain-specific information in the form of feature summarization, can be a promising direction to enhance the effectiveness of test-time adaptation.
|
| 28 |
+
|
| 29 |
+
# 2 Background
|
| 30 |
+
|
| 31 |
+
# 2.1 Related Work
|
| 32 |
+
|
| 33 |
+
Test-time Adaptation. Adapting machine learning models based on test samples has garnered growing interests in both generative problems such as super-resolution [13], image synthesis [14] and image manipulation [15], and discriminative problems like image classification [16]. Our work is focused on the latter one in the presence of distributional shifts. Several recent works [16, 17] have shown the advantage of adapting the learned classifier to new test domains in the unsupervised manner, without revisiting the source data. One simplest form is to replace the batch-norm statistics estimated on the training set with those on test examples [17]. Another line of work proposed to adapt the model parameters by exploiting the predicted labels on test examples, such as entropy minimization [8] and pseudo-labeling [7]. While these methods yield promising results on some benchmarks, they are inherently restricted to classification problems and often vulnerable under large distribution shifts [18].
|
| 34 |
+
|
| 35 |
+
More closely related to ours, [6] proposed test-time training through self-supervised learning, e.g., predicting the type of image rotations. This approach does not involve any assumptions about the output for the main task and is therefore more generic. It has been successfully applied to a variety of problems, such as instance tracking [19] and reinforcement learning [20]. Nevertheless, it was shown empirically inferior to other test-time algorithms [8]. Our work provides an in-depth analysis of its limitations, introduces simple yet effective remedies, and consolidates more theoretical grounds.
|
| 36 |
+
|
| 37 |
+
Feature Alignment. Aligning the distributions of training and test samples in the feature space is commonly used for domain adaptation. Previous feature alignment methods fall into two main categories: minimizing a divergence measure, such as MMD [21], Coral [22] and CMD [23], or encouraging the domain confusion through adversarial training [5, 24]. However, most of these methods rely on the co-existence of source and target data, and thus cannot be readily applied to the test-time setting where source data is not available. Our work revisits the critical role of feature alignment for test-time training and proposes a simple and practical strategy that enables online feature alignment, even with a limited batch size.
|
| 38 |
+
|
| 39 |
+

|
| 40 |
+
Figure 1: Illustration of a failure case where TTT hurts robustness under distributional shifts. (a) The predictive model reaches high accuracy on both the main classification task, i.e., separating red and blue data points, and the auxiliary self-supervised task, i.e., separating circles and crosses, in the training domain. (b) Given a large distributional shift, test samples are encoded into a new subspace, resulting in limited accuracy on both tasks. (c) Without any constraints on the feature distribution, TTT may result in an updated encoder severely overfitting to the SSL task, and consequently deteriorate the accuracy on the main task, as opposed to improving it.
|
| 41 |
+
|
| 42 |
+
Self-Supervised Learning. Self-supervised learning is a powerful paradigm to learn rich representations from unlabeled samples. Stunning progress has been made in recent years by designing informative self-supervised tasks [11, 12, 25–28] and stabilizing the training process [9, 29]. Prior works are mainly focused on unsupervised pre-training, whereas our work looks into the importance of incorporating strong self-supervised learning methods for test-time adaptation.
|
| 43 |
+
|
| 44 |
+
# 2.2 Preliminary: Test-Time Training
|
| 45 |
+
|
| 46 |
+
Test-time training (TTT) [6] is a general framework for adapting neural network models to a new test distribution based on unlabeled samples. Different from the conventional approach that trains the model only on the task of interest, TTT considers two tasks: a main task and an auxiliary SSL task. The model is trained on both tasks simultaneously with a multi-task architecture composed of one shared encoder $g$ and two separate heads $\pi _ { m }$ and $\pi _ { s }$ respectively. Given a labeled training dataset $D = \{ ( x _ { i } , y _ { i } ) \} _ { i \in \{ 1 , . . . , N \} }$ , the model is trained to minimize two losses jointly:
|
| 47 |
+
|
| 48 |
+
$$
|
| 49 |
+
\mathcal { L } _ { t r a i n } ( D ; g , \pi _ { m } , \pi _ { s } ) = \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \mathcal { L } _ { m } ( x _ { i } , y _ { i } ; g , \pi _ { m } ) + \lambda \mathcal { L } _ { s } ( x _ { i } ; g , \pi _ { s } ) ,
|
| 50 |
+
$$
|
| 51 |
+
|
| 52 |
+
where $\lambda$ is a hyper-parameter to balance the two tasks.
|
| 53 |
+
|
| 54 |
+
In the presence of distributional shifts, the learned model often struggles to directly generalize to a new test set $D ^ { \prime } = \{ x _ { i } ^ { \prime } \} _ { i \in \{ 1 , \dots , N ^ { \prime } \} }$ . The core idea of TTT is to fine-tune the encoder $g$ based on the self-supervised task with the test examples,
|
| 55 |
+
|
| 56 |
+
$$
|
| 57 |
+
\mathcal { L } _ { T T T } ( D ^ { \prime } ; g ^ { \prime } ) = \frac { 1 } { N ^ { \prime } } \sum _ { i = 1 } ^ { N ^ { \prime } } \mathcal { L } _ { s } ( x _ { i } ^ { \prime } ; g ^ { \prime } ) ,
|
| 58 |
+
$$
|
| 59 |
+
|
| 60 |
+
with the hope that the updated model $\pi _ { m } \circ g ^ { \prime }$ yields improved results on the main task.
|
| 61 |
+
|
| 62 |
+
TTT instantiated with a self-supervised rotation prediction task has been demonstrated effective for improving the robustness of image classifiers under common distributional shifts [6]. However, it was shown inferior to other families of test-time adaptation methods [7, 8]. We will next look into its strengths and limitations, and propose an improved version for better adaptation performance.
|
| 63 |
+
|
| 64 |
+
# 3 When Does Test-Time Training Fail?
|
| 65 |
+
|
| 66 |
+
In this section, we first throw light on a caveat of test-time training under large distributional shifts, and subsequently introduce practical solutions tailored for the test-time setting.
|
| 67 |
+
|
| 68 |
+

|
| 69 |
+
Figure 2: Our modified version of test-time training $\mathrm { ( T T T + + ) }$ ). Our method consists of three stages: model training, offline feature summarization, and online test-time adaptation. (i) During training, the model is optimized for the main task and an auxiliary contrastive self-supervised task jointly. (ii) Once training completes, we summarize the feature distributions after the encoder and the self-supervised head in the form of first and second-order moments. (iii) At test time, we adapt the encoder through online feature alignment (Sec. 3.2) and self-supervised learning (Sec. 4.2). In case of limited batch size, we maintain a large dynamic queue of feature vectors for robust moment estimates (Sec. 3.3).
|
| 70 |
+
|
| 71 |
+
# 3.1 Illustrative Example of Failures
|
| 72 |
+
|
| 73 |
+
One implicit assumption behind TTT is that the encoder update based on SSL can counter the effect of the underlying distributional shift on the main task. This assumption, however, is likely broken under large shifts and results in unexpected adaptation failures. To illustrate this limitation, we introduce a simple toy problem in Figure 1, where the main task and the SSL task are defined as classifying colors and symbols of encoded features in the 2-dimensional latent space. We consider an ideal scenario where the two tasks are highly correlated such that a well-trained model can attain high accuracies on both of them in the training domain. Nevertheless, in the presence of a significant distributional shift, the model may still suffer from substantial prediction errors in the test domain.
|
| 74 |
+
|
| 75 |
+
In this scenario, while TTT may restore the discriminative power of the learned representation for the main task to a certain degree, the unconstrained self-supervised adaptation may yield severe overfitting to the auxiliary SSL task. As a consequence, the performance on the main task can even deteriorate as opposed to improving. This phenomenon is not restricted to our illustration and also occurs in practice, as shown in Section 5.1.
|
| 76 |
+
|
| 77 |
+
# 3.2 Online Feature Alignment
|
| 78 |
+
|
| 79 |
+
As illustrated in the toy example above, simply applying self-supervised adaptation at test time can lead to arbitrarily poor results. To address this issue, we introduce an online feature alignment strategy to ensure robust adaptation at test time. The core idea of our strategy is to impose a constraint over the feature space during TTT such that the feature distribution of test examples remains close to that in the training domain. While some feature alignment techniques such as MMD [2] and adversarial training [24] have been widely used for domain adaptation, they often rely on sampling from training and test domains concurrently, which is impractical in the test-time setting. We, therefore, turn to classical divergence measures that can be estimated independently for each distribution. More specifically, we use the square distance of the first and second moments between two feature distributions, inspired by DDC [30] and Coral [22], to approximate the domain discrepancy. This design choice allows us to summarize the distribution of training features in a compact format and store it as part of the model, eliminating the need to revisit the training data during test-time adaptation.
|
| 80 |
+
|
| 81 |
+
Concretely, once training completes, we perform an offline feature summarization step that characterizes the disempirical mean ure vectors in the trainiand covariance matrix $Z = \{ z _ { 1 } ^ { T } , \dots , z _ { N } ^ { T } \}$ h the. The $\begin{array} { r } { \mu _ { z } = \frac { 1 } { N } \sum _ { i } ^ { N } z _ { i } } \end{array}$ $\begin{array} { r } { \Sigma _ { Z } = \frac { 1 } { N - 1 } \big ( Z ^ { T } Z - ( I ^ { T } Z ) ^ { T } ( I ^ { T } Z ) \big ) } \end{array}$ former is essentially equivalent to the channel-wise batch normalization statistics while the latter is more informative yet light-weight for computation and storage. At test time, we regularize the self-supervised adaptation by minimizing the distance between the feature statistics estimated from a mini-batch of test samples, i.e., $\mu _ { z } ^ { \prime }$ and $\Sigma _ { z } ^ { \prime }$ , and the pre-stored quantities about the training domain:
|
| 82 |
+
|
| 83 |
+
$$
|
| 84 |
+
\mathcal { L } _ { f , z } = \left\| \mu _ { z } - \mu _ { z } ^ { \prime } \right\| _ { 2 } ^ { 2 } + \left\| \Sigma _ { z } - \Sigma _ { z } ^ { \prime } \right\| _ { F } ^ { 2 } ,
|
| 85 |
+
$$
|
| 86 |
+
|
| 87 |
+
where $\lVert \cdot \rVert _ { 2 }$ is the Euclidean norm and $\left\| \cdot \right\| _ { F }$ is the Frobenius norm.
|
| 88 |
+
|
| 89 |
+
The basic form of online moment matching can be limiting in that the low-order statistics may be insufficient to fully capture complex distributions in high dimensions, e.g., 2048 for the standard ResNet-50. To alleviate this issue, we align the feature distributions at both the output of the encoder and the output of the self-supervised head, which are of lower dimensions, e.g., 128 in the case of contrastive learning described in Sec. 4.2. Our final objective at test time is a weighted combination of the self-supervised loss $\mathcal { L } _ { s }$ , the feature alignment loss at the encoder $\mathcal { L } _ { f , z }$ and that at the self-supervised head $\mathcal { L } _ { f , s }$ ,
|
| 90 |
+
|
| 91 |
+
$$
|
| 92 |
+
\mathcal { L } _ { T T T + + } = \mathcal { L } _ { s } + \lambda _ { z } \mathcal { L } _ { f , z } + \lambda _ { s } \mathcal { L } _ { f , s } ,
|
| 93 |
+
$$
|
| 94 |
+
|
| 95 |
+
where $\lambda _ { z }$ and $\lambda _ { s }$ are hyper-parameters controlling the emphasis on each term.
|
| 96 |
+
|
| 97 |
+
# 3.3 Online Dynamic Queue
|
| 98 |
+
|
| 99 |
+
One practical challenge for online feature alignment lies in scaling the strategy to problems having large numbers of classes. Intuitively, a good estimate of moments of the entire distribution needs at least a handful of samples per class. As a consequence, the demand for sample size grows linearly with the number of classes, for instance, over $\mathord { \sim } 1 0 0 0$ samples are required in the case of CIFAR-100. However, the computational resources during deployment are often limited to accommodate such a large batch size in the test-time setting.
|
| 100 |
+
|
| 101 |
+
To overcome this challenge, we draw inspiration from recent literature [9] and maintain a dynamic queue of encoded features to decouple the batch size from the sample size for moment estimates. More specifically, we construct a dynamic queue that contains a few batches of feature vectors encoded at test time. We progressively update the queue by appending the latest mini-batch and popping out the oldest one, as illustrated in Figure 2. This batch-queue decoupling allows us to collect a large and consistent pool of samples for online moment matching, even with a very limited batch size.
|
| 102 |
+
|
| 103 |
+
By integrating the online moment matching and batch-queue decoupling, our test-time algorithm takes into account both the discriminative power and the marginal distribution of the updated representations, enabling more robust adaptation under various settings. It is worth noting that the current moment matching strategy is just a particular instance of the online feature alignment scheme. It can be naturally extended to incorporate higher-order statistics [23, 31] to bring further performance gain at the cost of larger space and computational complexities.
|
| 104 |
+
|
| 105 |
+
# 4 When Does Test-Time Training Thrive?
|
| 106 |
+
|
| 107 |
+
Given properly aligned feature distribution, we next look into the potential of test-time training given strong SSL tasks. We first derive a lower bound of the test accuracy in general scenarios and then analyze it in a specific setup where the performance on the main task can be directly estimated. These analyses motivate our use of more informative self-supervised learning for test-time training.
|
| 108 |
+
|
| 109 |
+
# 4.1 Theoretical Results
|
| 110 |
+
|
| 111 |
+
We consider a training set comprised of samples drawn from the joint distribution $\mathbb { P } _ { X , Y _ { m } , Y _ { s } }$ , where $X , Y _ { m }$ and $Y _ { s }$ are random variables corresponding to the training samples, the main task labels and the self-supervised labels respectively. Similarly, the test set consists of samples drawn from the joint distribution $\mathbb { P } _ { X ^ { \prime } , Y _ { m } ^ { \prime } , Y _ { s } ^ { \prime } }$ . In the presence of distribution shift, $\mathbb { P } _ { X , Y _ { m } , Y _ { s } }$ and $\mathbb { P } _ { X ^ { \prime } , Y _ { m } ^ { \prime } , Y _ { s } ^ { \prime } }$ are not identical. However, we make the following assumption about label distribution in our analysis.
|
| 112 |
+
|
| 113 |
+
Assumption 1. The training and test labels are equal in distribution, $Y _ { m } \ { \overset { d } { = } } \ Y _ { m } ^ { \prime }$ , $Y _ { s } \ { \overset { d } { = } } \ Y _ { s } ^ { \prime }$ and $( Y _ { m } , Y _ { s } ) \overset { d } { = } ( Y _ { m } ^ { \prime } , Y _ { s } ^ { \prime } ) .$ .
|
| 114 |
+
|
| 115 |
+
In addition, we restrict our analysis to the scenarios where both two tasks can be solved perfectly during training.
|
| 116 |
+
|
| 117 |
+
Assumption 2. There exist an encoder g and classifiers $\pi _ { m }$ and $\pi _ { s }$ such that $\mathbb { P } ( \pi _ { m } ( g ( X ) ) = Y _ { m } ) =$ 1 and $\bar { \mathbb { P } } ( \pi _ { s } ( g ( X ) ) = Y _ { s } ) = 1$ .
|
| 118 |
+
|
| 119 |
+
During TTT, the shared encoder $g$ is updated to $g ^ { \prime }$ such that the self-supervised head $\pi _ { s }$ fits the test data. Given our proposed online feature alignment in Equation 4, we assume the encoded features in the training and test domains, i.e., $Z$ and $Z ^ { \prime }$ , have the following property.
|
| 120 |
+
|
| 121 |
+
Assumption 3. The marginal feature distribution and the conditional feature distributions at test time are aligned with their counterparts during training, that is, $Z { \overset { d } { = } } Z ^ { \prime }$ and $( Z \mid Y _ { s } = k ) \overset { d } { = } ( Z ^ { \prime } \mid$ $Y _ { s } ^ { \prime } = k$ ) for all classes $k$ in the SSL task.
|
| 122 |
+
|
| 123 |
+
nder these assumptions, we consider the outcome of test-time training in the worst-case scena
|
| 124 |
+
|
| 125 |
+
Theorem 1. The prediction accuracy on the main task is lower bounded:
|
| 126 |
+
|
| 127 |
+
$$
|
| 128 |
+
{ \mathbb { P } } ( \pi _ { m } ( Z ^ { \prime } ) = Y _ { m } ^ { \prime } ) \geq \sum _ { y _ { s } } { \mathbb { P } } ( Y _ { s } = y _ { s } ) \operatorname* { m a x } \left\{ 0 , 2 \left( \operatorname* { m a x } _ { y _ { m } } { \mathbb { P } } ( Y _ { m } = y _ { m } \mid Y _ { s } = y _ { s } ) - 0 . 5 \right) \right\} .
|
| 129 |
+
$$
|
| 130 |
+
|
| 131 |
+
Proof. Please refer to Section A.1 in the supplementary material.
|
| 132 |
+
|
| 133 |
+
Theorem 1 highlights the importance of the relation between the main task and the SSL task for the test accuracy after adaptation. More specifically, the SSL task needs to be informative with respect to the main task to guarantee the performance, i.e., knowing the SSL class $y _ { s }$ makes a main class $y _ { m }$ highly probable, or equivalently $\mathbb { P } ( Y _ { m } = y _ { m } \mid Y _ { s } = y _ { s } )$ is large, leading to a greater lower bound.
|
| 134 |
+
|
| 135 |
+
To further understand the impact of the task relation on test-time training, we next consider a particular setting, where the encoded features fully overfit to the SSL task (e.g., lengthy test-time training) and become independent of the main task label given the SSL label. Under this condition, the prediction accuracy on the main task can be directly estimated as follows.
|
| 136 |
+
|
| 137 |
+
Theorem 2. If $Z ^ { \prime } \perp \perp Y _ { m } ^ { \prime } \mid Y _ { s } ^ { \prime }$ , then the prediction accuracy on the main task is given by
|
| 138 |
+
|
| 139 |
+
$$
|
| 140 |
+
\mathbb { P } \big ( \pi _ { m } ( Z ^ { \prime } ) = Y _ { m } ^ { \prime } \big ) = \sum _ { y _ { s } } \left[ \mathbb { P } \big ( Y _ { s } = y _ { s } \big ) \sum _ { y _ { m } } \mathbb { P } \big ( Y _ { m } = y _ { m } \mid Y _ { s } = y _ { s } \big ) ^ { 2 } \right] .
|
| 141 |
+
$$
|
| 142 |
+
|
| 143 |
+
Proof. Please refer to Section A.2 in the supplementary material.
|
| 144 |
+
|
| 145 |
+
Intuitively, if the encoded features do not contain more information than the SSL labels about the main task, the accuracy of the main classifier $\pi _ { m }$ only depends on the property of the SSL task. In particular, the square term on the right-hand side of Equation 6 emphasizes the paramount importance of designing a closely related SSL task. When the two tasks diverge, the test accuracy drops quadratically fast, leading to ineffective adaptation.
|
| 146 |
+
|
| 147 |
+
# 4.2 Test-Time Training through Contrastive Learning
|
| 148 |
+
|
| 149 |
+
Our theoretical analysis above reveals the importance of incorporating an SSL task highly correlated with the main task for test-time training. One practical way to quantify the relation between two tasks is to measure the transferability of the representation learned from one task to another [32]. Given the remarkable results of contrastive methods for visual representation pre-training [9, 12, 33, 34], we hypothesize that they would also be suitable choices for test-time training. We thus replace the rotation prediction task with SimCLR [12] in the context of visual recognition. Given a mini-batch of $B$ images, we augment each image to two views. We consider the two augmented views from the same original instance as a positive pair and treat the other pairs as negative ones. The feature vector $z _ { i } = g ( \bar { x } _ { i } )$ of each image $x _ { i }$ is projected to a lower-dimensional space $h _ { i } = \pi _ { s } ( z _ { i } )$ through our self-supervised head. The projected hidden embeddings from a positive pair $< h _ { i } , h _ { j } >$ are encouraged to be closer than those from the negative ones through the following loss,
|
| 150 |
+
|
| 151 |
+
$$
|
| 152 |
+
\mathcal { L } _ { s } = - \log \frac { \exp ( \sin ( h _ { i } , h _ { j } ) / \tau ) } { \sum _ { k = 1 } ^ { 2 B } \mathbb { 1 } _ { k \neq i } \exp ( \sin ( h _ { i } , h _ { k } ) / \tau ) } ,
|
| 153 |
+
$$
|
| 154 |
+
|
| 155 |
+
where $\tau$ is a temperature scaling parameter. The distance between projected embeddings is measured by cosine similarity $\mathrm { s i m } ( u , v ) \stackrel { \smile } { = } u ^ { T } v / ( \| u \| \| v \| )$ .
|
| 156 |
+
|
| 157 |
+

|
| 158 |
+
Figure 3: Qualitative comparison of TTT [6] and our $\mathrm { T T } \mathrm { + + }$ on the inter-twinning moons problem in the presence of large translation and rotation shifts. The vanilla TTT drives the decision boundary further away from a desired one due to a severe feature misalignment, marked by the dashed arrow in the feature PCA visualization. In comparison, our method adapts the decision boundary to the test domain more effectively.
|
| 159 |
+
|
| 160 |
+

|
| 161 |
+
Figure 4: Quantitative comparison of TTT [6] and our $\mathrm { T T } \mathrm { + + }$ on the inter-twinning moons problem under 150 different setups. Standard deviations are visualized in shaded regions. Our method (left) is particularly advantageous under large shifts, i.e., low test accuracy before adaptation, and (right) greatly benefits from a higher correlation, i.e., larger label agreement, between the main and SSL tasks.
|
| 162 |
+
|
| 163 |
+
# 5 Experiments
|
| 164 |
+
|
| 165 |
+
We empirically validate our proposed method in four scenarios: synthetic toy problem, common image corruptions, natural domain shifts, and sim-to-real transfer.
|
| 166 |
+
|
| 167 |
+
We consider the following baselines throughout our experiments:
|
| 168 |
+
|
| 169 |
+
• Test: the model in the training domain is directly evaluated on the test data without any adaptation;
|
| 170 |
+
• Test-time normalization (BN) [35] updates the batch normalization statistics of the trained network according to the test data;
|
| 171 |
+
• Test-time entropy minimization (TENT) [8] updates the batch normalization statistics of the trained network by minimizing the entropy of the model predictions on the test data;
|
| 172 |
+
• Source Hypothesis Transfer (SHOT) [36] freezes the classifier module and updates only the feature extraction module by exploiting the concepts of information maximization and self-supervised pseudo-labeling during testing;
|
| 173 |
+
• Test-time training (TTT-R) [6] trains the network jointly on the main task and a rotation-based SSL task in the source domain; during test, TTT-R continues to train on the rotation-based task in the target domain.
|
| 174 |
+
|
| 175 |
+
We also evaluate the following ablated versions of our method:
|
| 176 |
+
|
| 177 |
+
• Test-time feature alignment (TFA) aligns the first-order and second-order statistics of the source and target distributions during testing (Section 3.2 and 3.3);
|
| 178 |
+
• Test-time contrastive learning (TTT-C) trains the network jointly on the main task and a contrastive learning (SSL) task in the source domain; during test, TTT-C continues to update the encoder based on contrastive learning in the target domain (Section 4.2).
|
| 179 |
+
|
| 180 |
+
The full version of our method, improved test-time training $\mathbf { \left( T T + \right) }$ ), combines TFA and TTT-C.
|
| 181 |
+
|
| 182 |
+
# 5.1 Synthetic Toy Problem
|
| 183 |
+
|
| 184 |
+
We first evaluate our method on the inter-twinning moons problem [4, 37], where the main task is to predict the moon class of a given data point and the SSL task is to predict on which side of the hyperplane (i.e., linear separator between the two moons) the data point lies on. The relation (label agreement) between the two tasks depends on the separation distance between the two moons. To solve both tasks simultaneously, we build a small neural network that consists of a 2-layer MLP as the shared encoder and two 2-layer MLPs as separate task heads. Each hidden layer contains 8 neurons. The learned model attains over $9 9 \%$ accuracy on both the main and SSL tasks in the training domain. We simulate a variety of distributional shifts through translation and rotation of all data points.
|
| 185 |
+
|
| 186 |
+
Figure 3 shows the decision boundaries and encoded features from the vanilla TTT and our proposed $\mathrm { T T T } { + } { + }$ in a particular test case of large distributional shift. TTT not only fails to improve the classification accuracy but even pushes the decision boundary further away from a desired one due to the severe feature distribution mismatch, as evidenced in the PCA visualization. In comparison, our $\mathrm { T T T } { + } { + }$ yields substantial performance gain, boosting the test accuracy from $50 \%$ to $93 \%$ , thanks to the reduced feature distributional shift enforced by the proposed online moment matching.
|
| 187 |
+
|
| 188 |
+

|
| 189 |
+
Figure 5: Classification error $( \% )$ on CIFAR10-C [1].
|
| 190 |
+
|
| 191 |
+
Table 1: Average classification error $( \% )$ on CIFAR10-C/100-C [1] and CIFAR10.1 [40]
|
| 192 |
+
|
| 193 |
+
<table><tr><td>Method</td><td>C10-C</td><td>C100-C</td><td>C10.1</td></tr><tr><td>Test</td><td>29.1</td><td>61.2</td><td>12.1</td></tr><tr><td>BN [41]</td><td>15.7</td><td>43.3</td><td>14.1</td></tr><tr><td>TTT-R [6]</td><td>14.3</td><td>40.4</td><td>11.0</td></tr><tr><td>SHOT [36]</td><td>14.7</td><td>38.1</td><td>11.1</td></tr><tr><td>TENT [8]</td><td>12.6</td><td>36.3</td><td>13.4</td></tr><tr><td>TFA (Ours)</td><td>11.9</td><td>35.8</td><td>12.1</td></tr><tr><td>TTT-C (Ours)</td><td>10.7</td><td>36.9</td><td>9.7</td></tr><tr><td>TTT++ (Ours)</td><td>9.8</td><td> 34.1</td><td>9.5</td></tr></table>
|
| 194 |
+
|
| 195 |
+
Figure 4 summarizes the quantitative results of our methods as well as the vanilla counterpart under 150 different simulated setups. As shown on the left graph, when the domain shift only causes mild test errors on the main task, both the original TTT and our $\mathrm { T T T } { \cdot } + +$ yield strong adaptation results. However, the effectiveness of TTT deteriorates quickly along with the growth of domain shift, as reflected on the lower test accuracy. In comparison, our proposed $\mathrm { T T T } { + } { + }$ demonstrates clear advantages under large shifts, e.g., when the test accuracy before adaptation is around 0.5. We further examine the impact of the relation between the main and SSL tasks on the adaptation performance by varying the separation distance between the two moons. The simulation results confirm the high potential of test-time training given improved SSL tasks, as analyzed in Section 4.1.
|
| 196 |
+
|
| 197 |
+
# 5.2 Common Image Corruption
|
| 198 |
+
|
| 199 |
+
We further assess the robustness of our method against common image corruptions. Following the evaluation protocol of previous work [8], we train ResNet-50 [38] on CIFAR10/CIFAR100 [39] and test it on the CIFAR10-C/CIFAR100-C [1] datasets, which contain 15 types of algorithmically generated corruptions, such as noise, blur and snow effects. We use a batch size of 256 for test-time adaptation. In addition, we use a dynamic queue containing 16 batches of feature vectors for online feature alignment on CIFAR100-C.
|
| 200 |
+
|
| 201 |
+
Figure 5 shows the quantitative results under each type of image corruption. The average results on CIFAR10-C and CIFAR100-C are reported in Table 1. Our $\mathrm { T T T } { + } { + }$ clearly outperforms the prior state-of-the-art test-time methods on CIFAR10-C and CIFAR100-C. In particular, incorporating a strong self-supervision task (TTT-C) already suffices to perform on par or better than TENT. Adding test-time feature alignment (TFA) on top of that yields an additional $\sim 8 \%$ relative reduction in terms of the test error.
|
| 202 |
+
|
| 203 |
+
# 5.3 Natural Domain Shift
|
| 204 |
+
|
| 205 |
+
We next demonstrate the efficacy of our method $\mathrm { T T T } { + } { + }$ to tackle natural distribution shifts. We again use the pre-trained ResNet-50 and test it on CIFAR10.1 [40], a recently collected test set subject to natural distributional shift. Despite its high perceptual similarity with the CIFAR10 dataset, the CIFAR10.1 typically leads to a drop of accuracy $4 \%$ to $10 \%$ ) for a wide range of deep models [40].
|
| 206 |
+
|
| 207 |
+
Table 1 summarizes the results of various test-time algorithms on CIFAR10.1. Previous batch-normbased methods perform poorly and even degrade model accuracy. This phenomenon is tied to their implicit assumption that different samples and spatial locations are shifted in a similar manner [18], which is true for algorithmically generated image corruptions but does not hold on CIFAR-10.1. In contrast, $\mathrm { T T T } { + } { + }$ is more generic and yields stronger performance under the natural distribution shift.
|
| 208 |
+
|
| 209 |
+
# 5.4 Sim-to-Real Transfer
|
| 210 |
+
|
| 211 |
+
We finally validate the effectiveness of our method on the VisDA-C dataset [42], a challenging large-scale benchmark of synthetic-to-real object classification. As shown in Table 2, prior methods that are fairly competitive under image corruptions, such as BN [41] and TENT [8], are not effective on VisDA-C. We conjecture that this is attributed to their strong restrictions over the adaptable parameters at test time. In contrast, our proposed method is more flexible, allows the model to update the entire encoder, and thus achieves compelling results on VisDA-C. Furthermore, the test-time feature alignment plays a crucial role in this synthetic-to-real domain adaptation problem, providing $\sim 1 3 \%$ performance boost on top of the TTT-C.
|
| 212 |
+
|
| 213 |
+
Table 2: Classification error $( \% )$ on the large-scale VisDA-C dataset [42].
|
| 214 |
+
|
| 215 |
+
<table><tr><td>Method</td><td>plane</td><td>bcycl</td><td>bus</td><td>car</td><td>horse</td><td>knife</td><td>mcycl</td><td>person</td><td>plant</td><td>sktbrd</td><td>train</td><td>truck</td><td>Per-class</td></tr><tr><td>Test</td><td>56.52</td><td>88.71</td><td>62.77</td><td>30.56</td><td>81.88</td><td>99.03</td><td>17.53</td><td>95.85</td><td>51.66</td><td>77.86</td><td>20.44</td><td>99.51</td><td>58.72</td></tr><tr><td>BN [41]</td><td>44.38</td><td>56.98</td><td>33.24</td><td>55.28</td><td>37.45</td><td>66.60</td><td>16.55</td><td>59.02</td><td>43.55</td><td>60.72</td><td>31.07</td><td>82.98</td><td>48.12</td></tr><tr><td>TENT [8]</td><td>13.43</td><td>77.98</td><td>20.17</td><td>48.15</td><td>21.72</td><td>82.45</td><td>12.37</td><td>35.78</td><td>21.06</td><td>76.41</td><td>34.11</td><td>98.93</td><td>42.73</td></tr><tr><td>SHOT [36]</td><td>5.73</td><td>13.64</td><td>23.33</td><td>42.69</td><td>7.93</td><td>86.99</td><td>19.17</td><td>19.97</td><td>11.63</td><td>11.09</td><td>15.06</td><td>43.26</td><td>25.04</td></tr><tr><td>TFA (Ours)</td><td>28.25</td><td>32.03</td><td>33.67</td><td>64.77</td><td>20.49</td><td>56.63</td><td>22.52</td><td>36.30</td><td>24.84</td><td>35.20</td><td>25.31</td><td>64.24</td><td>39.58</td></tr><tr><td>TTT-C (Ours)</td><td>5.46</td><td>32.23</td><td>25.42</td><td>37.03</td><td>7.84</td><td>85.20</td><td>9.14</td><td>23.80</td><td>11.72</td><td>11.00</td><td>7.74</td><td>56.87</td><td>25.72</td></tr><tr><td>TTT++ (Ours)</td><td>4.13</td><td>26.20</td><td>21.60</td><td>31.70</td><td>7.43</td><td>83.30</td><td>7.83</td><td>21.10</td><td>7.03</td><td>7.73</td><td>6.91</td><td>51.40</td><td>22.46</td></tr></table>
|
| 216 |
+
|
| 217 |
+
Table 3: Classification error $( \% )$ results from online feature alignment with or without a dynamic queue of feature vectors on the CIFAR100-C under level-5 fog corruption. Sample size $=$ Batch size $\times \#$ Batches. Given a fixed batch size, enlarging the queue size leads to similar results as having a larger batch size.
|
| 218 |
+
|
| 219 |
+
<table><tr><td></td><td colspan="3">w/o queue</td><td colspan="4">w/ queue</td></tr><tr><td>Sample Size</td><td>64</td><td>128</td><td>256</td><td>64×2</td><td>64×4</td><td>64×8</td><td>64 ×16</td></tr><tr><td>Test Error</td><td>40.31</td><td>38.67</td><td>37.01</td><td>39.84</td><td>37.37</td><td>36.18</td><td>36.02</td></tr></table>
|
| 220 |
+
|
| 221 |
+
# 5.5 Effect of Batch-Queue Decoupling
|
| 222 |
+
|
| 223 |
+
To verify the effects of batch-queue decoupling, we compare the results of online feature alignment with different sample sizes. Table 3 summarizes the test errors on CIFAR100 under the level-5 fog corruption. As expected, larger sample sizes generally lead to lower classification errors. Interestingly, while the performance of using a dynamic queue is slightly worse than its counterpart of the same sample size from a single large batch, enlarging the queue size always yields stronger results. For instance, given a small batch size of 64, using a dynamic queue maintaining 512 or 1024 feature vectors from 8 or 16 consecutive batches respectively is more advantageous than the vanilla moment matching based on a batch size of 256 samples. This result corroborates the benefit of integrating a dynamic queue into our proposed feature alignment framework, for enhancing the scalability in the online setting.
|
| 224 |
+
|
| 225 |
+
# 5.6 Design Choice for Moment Matching
|
| 226 |
+
|
| 227 |
+
As discussed in Section 3, our proposed online feature alignment can be instantiated with different orders of moments and applied at various layers. To understand the effects of the detailed design choices, we empirically compare our proposed version against several ablated variants in Table 4. Irrespective of the layer choice, our proposed online feature alignment consistently results in reduced classification error. Nevertheless, moment matching applied to the self-supervised head alone leads to lower test error compared to applying it to the feature extractor output in 11 out of 15 types of corruption. We conjecture that the strong performance of the former one is attributed to the lower dimensionality of the feature vector, which allows for a more accurate estimate of feature statistics. The best result comes from the online feature alignment at the outputs of both the encoder and the projection head, which validates our design choice in Section 3.2.
|
| 228 |
+
|
| 229 |
+
We further validate the choice of divergence measure through an ablation study. The online feature alignment using the second-order moment (covariance) leads to clearly better results than the one using the first-order moment (mean). It is also evident that the online feature alignment is most effective when both the mean and the covariance are taken into account.
|
| 230 |
+
|
| 231 |
+
Table 4: Classification error $( \% )$ on CIFAR10-C [1] with different versions of online feature alignment. Taking into account both the first and second-order moments at two different layers is better than the other counterparts in terms of the robustness against most types of image corruptions.
|
| 232 |
+
|
| 233 |
+
<table><tr><td>TFA</td><td>brit</td><td>contr</td><td>defoc</td><td>elast</td><td>fog</td><td>frost</td><td>gauss</td><td>glass</td><td>impul</td><td>jpeg</td><td>motn</td><td>pixel</td><td>shot</td><td>snow</td><td>zoom</td></tr><tr><td>w/oLf.s</td><td>7.96</td><td>7.57</td><td>9.4</td><td>17.24</td><td>14.38</td><td>12.54</td><td>14.62</td><td>21.05</td><td>21.4</td><td>12.68</td><td>11.92</td><td>10.7</td><td>13.7</td><td>12.74</td><td>7.32</td></tr><tr><td>w/oLf,z</td><td>7.85</td><td>7.84</td><td>9.18</td><td>16.51</td><td>14.33</td><td>11.99</td><td>13.79</td><td>20.08</td><td>20.17</td><td>12.42</td><td>12.02</td><td>10.5</td><td>12.78</td><td>13.28</td><td>7.44</td></tr><tr><td>w/£</td><td>7.49</td><td>7.56</td><td>9.62</td><td>18.62</td><td>19.22</td><td>12.72</td><td>16.02</td><td>25.07</td><td>25.17</td><td>13.43</td><td>13.63</td><td>11.22</td><td>15.04</td><td>15.11</td><td>7.77</td></tr><tr><td>w/o μ</td><td>7.43</td><td>7.37</td><td>8.90</td><td>15.92</td><td>12.98</td><td>11.57</td><td>13.46</td><td>19.27</td><td>18.95</td><td>11.87</td><td>11.11</td><td>9.97</td><td>12.81</td><td>11.76</td><td>7.04</td></tr><tr><td>Full</td><td>7.44</td><td>7.40</td><td>8.89</td><td>15.73</td><td>12.82</td><td>11.49</td><td>12.94</td><td>18.46</td><td>19.13</td><td>11.66</td><td>10.77</td><td>9.93</td><td>12.67</td><td>11.73</td><td>7.03</td></tr></table>
|
| 234 |
+
|
| 235 |
+
# 6 Conclusion and Discussions
|
| 236 |
+
|
| 237 |
+
In this work, we conduct an in-depth analysis of the limitations and potential of self-supervised test-time training. We draw attention to the risk of feature distribution mismatch, which is critical but largely overlooked in recent test-time algorithms. We shed light on the strong potential of this approach by analyzing the growth of test accuracy given improved SSL tasks. These analyses inspire three proposed modifications, namely online feature alignment, batch-queue decoupling and contrastive test-time training, which yield state-of-the-art results on multiple robustness benchmarks. Our results suggest the advantages of bringing additional task-specific and domain-specific information in a compact format for test-time adaptation. We hope these findings will motivate researchers and practitioners to rethink what should be stored, in addition to weight parameters, for the robust deployment of machine learning models.
|
| 238 |
+
|
| 239 |
+
Limitations. In this work, we restrain feature summarization to first and second-order moments. Yet, the low-order statistics may be insufficient to characterize the complex distribution of highdimensional features. Developing more advanced summarization methods tailored for test-time adaptation is an interesting avenue for future work. In addition, there may exist a considerable gap between our theoretical analysis and the empirical results, when the stated assumptions do not hold. For instance, neither the classification heads nor the feature alignment is perfect in practice. More theoretical guarantees can be valuable for the practical use of test-time adaptation in safety-critical scenarios.
|
| 240 |
+
|
| 241 |
+
Open Questions. In our experiments, we only consider the standard ResNet-50 as the backbone architecture and share the whole feature extractor between the main task and self-supervised task. Yet, recent literature [30, 43] has shown that different layers capture different levels of semantic granularity. The impact of architectural design on test-time training remains an open question. Furthermore, while we empirically compare our proposed method against other families of test-time adaptation algorithms, these techniques exploit different supervisory signals extracted from unlabeled data, which can be complementary to each other. Blending these techniques into a unified framework is another interesting direction to explore in the future.
|
| 242 |
+
|
| 243 |
+
Societal Impact. Our work aims at expanding the current horizon of machine learning algorithms for test-time adaptation. For applications where humans’ lives are at risk, such as autonomous driving, trust, safety, robustness are all mandatory keywords. The field has made amazing progress when the training and testing environments are highly similar. What happens when a machine gets deployed in a new environment? We, humans, have an innate capability for handling such shifts. We believe that machines should have the same capability as humans. Indeed, there is a long way to go. Nevertheless, we hope that our work will foster more research in analyzing and devising algorithms for robust and effective adaptation at test time.
|
| 244 |
+
|
| 245 |
+
# Acknowledgements
|
| 246 |
+
|
| 247 |
+
This work was supported by the Swiss National Science Foundation under the Grant 2OOO21- L92326, Honda R&D Co. Ltd, EPFL Open Science fund and Valeo. We thank Sudeep Salgia, Tao Lin, Lingjun Meng, Yifan Sun for helpful inputs to our early drafts and anonymous reviewers for valuable comments.
|
| 248 |
+
|
| 249 |
+
References
|
| 250 |
+
[1] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019.
|
| 251 |
+
[2] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning Transferable Features with Deep Adaptation Networks. In International Conference on Machine Learning, pages 97–105. PMLR, June 2015. ISSN: 1938-7228.
|
| 252 |
+
[3] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I. Jordan. Deep Transfer Learning with Joint Adaptation Networks. In Proceedings of the 34th International Conference on Machine Learning, pages 2208–2217. PMLR, July 2017. ISSN: 2640-3498.
|
| 253 |
+
[4] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The Journal of Machine Learning Research, 17(1):2096–2030, January 2016.
|
| 254 |
+
[5] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial Discriminative Domain Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7167–7176, 2017.
|
| 255 |
+
[6] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229–9248. PMLR, November 2020. ISSN: 2640-3498.
|
| 256 |
+
[7] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning (ICML), pages 6028–6039, July 13–18 2020.
|
| 257 |
+
[8] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully Test-Time Adaptation by Entropy Minimization. In International Conference on Learning Representations, September 2020.
|
| 258 |
+
[9] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729–9738, 2020.
|
| 259 |
+
[10] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality Reduction by Learning an Invariant Mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR’06), volume 2, pages 1735–1742, New York, NY, USA, 2006. IEEE.
|
| 260 |
+
[11] Zhirong Wu, Yuanjun Xiong, Stella Yu, and Dahua Lin. Unsupervised Feature Learning via Non-parametric Instance Discrimination. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3733–3742, June 2018. ISSN: 2575-7075.
|
| 261 |
+
[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 1597–1607. PMLR, November 2020. ISSN: 2640- 3498.
|
| 262 |
+
[13] Assaf Shocher, Nadav Cohen, and Michal Irani. Zero-Shot Super-Resolution Using Deep Internal Learning. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3118–3126, June 2018. ISSN: 2575-7075.
|
| 263 |
+
[14] Yuejiang Liu, Parth Kothari, and Alexandre Alahi. Collaborative Sampling in Generative Adversarial Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):4948– 4956, April 2020. Number: 04.
|
| 264 |
+
[15] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM Transactions on Graphics, 38(4):59:1–59:11, July 2019.
|
| 265 |
+
|
| 266 |
+
[16] Jogendra Nath Kundu, Naveen Venkat, Rahul M. V, and R. Venkatesh Babu. Universal SourceFree Domain Adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544–4553, 2020.
|
| 267 |
+
|
| 268 |
+
[17] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33, 2020.
|
| 269 |
+
|
| 270 |
+
[18] Collin Burns and Jacob Steinhardt. Limitations of Post-Hoc Feature Alignment for Robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2525–2533, 2021.
|
| 271 |
+
|
| 272 |
+
[19] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to Track Instances without Video Annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8680–8689, 2021.
|
| 273 |
+
|
| 274 |
+
[20] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A. Efros, Lerrel Pinto, and Xiaolong Wang. Self-Supervised Policy Adaptation during Deployment. In International Conference on Learning Representations, September 2020.
|
| 275 |
+
|
| 276 |
+
[21] Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander Smola. A Kernel Two-Sample Test. Journal of Machine Learning Research, 13(25):723–773, 2012.
|
| 277 |
+
|
| 278 |
+
[22] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation Alignment for Unsupervised Domain Adaptation. In Gabriela Csurka, editor, Domain Adaptation in Computer Vision Applications, Advances in Computer Vision and Pattern Recognition, pages 153–171. Springer International Publishing, Cham, 2017.
|
| 279 |
+
|
| 280 |
+
[23] Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschläger, and Susanne Saminger-Platz. Central Moment Discrepancy (CMD) for Domain-Invariant Representation Learning. International Conference on Learning Representations, International Conference on Learning Representations, November 2016. 00283.
|
| 281 |
+
|
| 282 |
+
[24] Yaroslav Ganin and Victor Lempitsky. Unsupervised Domain Adaptation by Backpropagation. In Proceedings of the 32nd International Conference on Machine Learning, pages 1180–1189. PMLR, June 2015. ISSN: 1938-7228.
|
| 283 |
+
|
| 284 |
+
[25] Mehdi Noroozi and Paolo Favaro. Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, European Conference on Computer Vision, pages 69–84, Cham, 2016. Springer International Publishing.
|
| 285 |
+
|
| 286 |
+
[26] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep Clustering for Unsupervised Learning of Visual Features. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, Lecture Notes in Computer Science, pages 139–156, Cham, 2018. Springer International Publishing.
|
| 287 |
+
|
| 288 |
+
[27] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social NCE: Contrastive Learning of SociallyAware Motion Representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15118–15129, 2021.
|
| 289 |
+
|
| 290 |
+
[28] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021.
|
| 291 |
+
|
| 292 |
+
[29] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750–15758, 2021.
|
| 293 |
+
|
| 294 |
+
[30] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell. Deep Domain Confusion: Maximizing for Domain Invariance. arXiv:1412.3474 [cs], December 2014. arXiv: 1412.3474.
|
| 295 |
+
|
| 296 |
+
[31] Chao Chen, Zhihang Fu, Zhihong Chen, Sheng Jin, Zhaowei Cheng, Xinyu Jin, and Xiansheng Hua. HoMM: Higher-Order Moment Matching for Unsupervised Domain Adaptation. Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):3422–3429, April 2020. Number: 04.
|
| 297 |
+
|
| 298 |
+
[32] Amir R. Zamir, Alexander Sax, William Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling Task Transfer Learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3712–3722, 2018.
|
| 299 |
+
|
| 300 |
+
[33] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised Learning. arXiv:2006.07733 [cs, stat], September 2020. 00623 arXiv: 2006.07733.
|
| 301 |
+
|
| 302 |
+
[34] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. arXiv:2006.09882 [cs], January 2021. 00403 arXiv: 2006.09882.
|
| 303 |
+
|
| 304 |
+
[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Removing covariate shift improves robustness against common corruptions. CoRR, abs/2006.16971, 2020.
|
| 305 |
+
|
| 306 |
+
[36] Jian Liang, Dapeng Hu, and Jiashi Feng. Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation. In International Conference on Machine Learning, pages 6028–6039. PMLR, November 2020. ISSN: 2640-3498.
|
| 307 |
+
|
| 308 |
+
[37] Pascal Germain, Amaury Habrard, François Laviolette, and Emilie Morvant. A PAC-Bayesian Approach for Domain Adaptation with Specialization to Linear Classifiers. In Proceedings of the 30th International Conference on Machine Learning, pages 738–746. PMLR, May 2013. ISSN: 1938-7228.
|
| 309 |
+
|
| 310 |
+
[38] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016.
|
| 311 |
+
|
| 312 |
+
[39] A. Krizhevsky. Learning multiple layers of features from tiny images, 2009.
|
| 313 |
+
|
| 314 |
+
[40] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet Classifiers Generalize to ImageNet? In Proceedings of the 36th International Conference on Machine Learning, pages 5389–5400. PMLR, May 2019. ISSN: 2640-3498.
|
| 315 |
+
|
| 316 |
+
[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 448–456. JMLR.org, 2015.
|
| 317 |
+
|
| 318 |
+
[42] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. ArXiv, abs/1710.06924, 2017.
|
| 319 |
+
|
| 320 |
+
[43] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How Transferable Are Features in Deep Neural Networks? In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, NIPS’14, pages 3320–3328, Cambridge, MA, USA, 2014. MIT Press.
|
| 321 |
+
|
| 322 |
+
[44] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised Contrastive Learning. arXiv:2004.11362 [cs, stat], April 2020. arXiv: 2004.11362.
|
parse/train/86NHK__yFDl/86NHK__yFDl_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/HJcSzz-CZ/HJcSzz-CZ.md
ADDED
|
@@ -0,0 +1,319 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# META-LEARNING FOR SEMI-SUPERVISED FEW-SHOTCLASSIFICATION
|
| 2 |
+
|
| 3 |
+
Mengye $\mathbf { R e n } ^ { \dagger \ltimes }$ , Eleni Triantafillou∗ †on, Sachin Ravi∗ §, Jake Snell†on, Kevin Swersky¶, Joshua B. Tenenbaum\, Hugo Larochelle¶‡ & Richard S. Zemel†‡on
|
| 4 |
+
|
| 5 |
+
†University of Toronto, §Princeton University, ¶Google Brain, \MIT, ‡CIFAR, onVector Institute {mren,eleni}@cs.toronto.edu, sachinr@cs.princeton.edu, jsnell@cs.toronto.edu, kswersky@google.com, jbt@mit.edu, hugolarochelle@google.com, zemel@cs.toronto.edu
|
| 6 |
+
|
| 7 |
+
# ABSTRACT
|
| 8 |
+
|
| 9 |
+
In few-shot classification, we are interested in learning algorithms that train a classifier from only a handful of labeled examples. Recent progress in few-shot classification has featured meta-learning, in which a parameterized model for a learning algorithm is defined and trained on episodes representing different classification problems, each with a small labeled training set and its corresponding test set. In this work, we advance this few-shot classification paradigm towards a scenario where unlabeled examples are also available within each episode. We consider two situations: one where all unlabeled examples are assumed to belong to the same set of classes as the labeled examples of the episode, as well as the more challenging situation where examples from other distractor classes are also provided. To address this paradigm, we propose novel extensions of Prototypical Networks (Snell et al., 2017) that are augmented with the ability to use unlabeled examples when producing prototypes. These models are trained in an end-to-end way on episodes, to learn to leverage the unlabeled examples successfully. We evaluate these methods on versions of the Omniglot and miniImageNet benchmarks, adapted to this new framework augmented with unlabeled examples. We also propose a new split of ImageNet, consisting of a large set of classes, with a hierarchical structure. Our experiments confirm that our Prototypical Networks can learn to improve their predictions due to unlabeled examples, much like a semi-supervised algorithm would.
|
| 10 |
+
|
| 11 |
+
# 1 INTRODUCTION
|
| 12 |
+
|
| 13 |
+
The availability of large quantities of labeled data has enabled deep learning methods to achieve impressive breakthroughs in several tasks related to artificial intelligence, such as speech recognition, object recognition and machine translation. However, current deep learning approaches struggle in tackling problems for which labeled data are scarce. Specifically, while current methods excel at tackling a single problem with lots of labeled data, methods that can simultaneously solve a large variety of problems that each have only a few labels are lacking. Humans on the other hand are readily able to rapidly learn new classes, such as new types of fruit when we visit a tropical country. This significant gap between human and machine learning provides fertile ground for deep learning developments.
|
| 14 |
+
|
| 15 |
+
For this reason, recently there has been an increasing body of work on few-shot learning, which considers the design of learning algorithms that specifically allow for better generalization on problems with small labeled training sets. Here we focus on the case of few-shot classification, where the given classification problem is assumed to contain only a handful of labeled examples per class. One approach to few-shot learning follows a form of meta-learning 1 (Thrun, 1998; Hochreiter et al., 2001), which performs transfer learning from a pool of various classification problems generated from large quantities of available labeled data, to new classification problems from classes unseen at training time. Meta-learning may take the form of learning a shared metric (Vinyals et al., 2016; Snell et al., 2017), a common initialization for few-shot classifiers (Ravi & Larochelle, 2017; Finn et al., 2017) or a generic inference network (Santoro et al., 2016; Mishra et al., 2017).
|
| 16 |
+
|
| 17 |
+
These various meta-learning formulations have led to significant progress recently in few-shot classification. However, this progress has been limited in the setup of each few-shot learning episode, which differs from how humans learn new concepts in many dimensions. In this paper we aim to generalize the setup in two ways. First, we consider a scenario where the new classes are learned in the presence of additional unlabeled data. While there have been many successful applications of semisupervised learning to the regular setting of a single classification task (Chapelle et al., 2010) where classes at training and test time are the same, such work has not addressed the challenge of performing transfer to new classes never seen at training time, which we consider here. Second, we consider the situation where the new classes to be learned are not viewed in isolation. Instead, many of the unlabeled examples are from different classes; the presence of such distractor classes introduces an additional and more realistic level of difficulty to the fewshot problem.
|
| 18 |
+
|
| 19 |
+

|
| 20 |
+
Figure 1: Consider a setup where the aim is to learn a classifier to distinguish between two previously unseen classes, goldfish and shark, given not only labeled examples of these two classes, but also a larger pool of unlabeled examples, some of which may belong to one of these two classes of interest. In this work we aim to move a step closer to this more natural learning framework by incorporating in our learning episodes unlabeled data from the classes we aim to learn representations for (shown with dashed red borders) as well as from distractor classes .
|
| 21 |
+
|
| 22 |
+
This work is a first study of this challenging semi-supervised form of few-shot learning. First, we define the problem and propose benchmarks for evaluation that are adapted from the Omniglot and miniImageNet benchmarks used in ordinary few-shot learning. We perform an extensive empirical investigation of the two settings mentioned above, with and without distractor classes. Second, we propose and study three novel extensions of Prototypical Networks (Snell et al., 2017), a state-ofthe-art approach to few-shot learning, to the semi-supervised setting. Finally, we demonstrate in our experiments that our semi-supervised variants successfully learn to leverage unlabeled examples and outperform purely supervised Prototypical Networks.
|
| 23 |
+
|
| 24 |
+
# 2 BACKGROUND
|
| 25 |
+
|
| 26 |
+
We start by defining precisely the current paradigm for few-shot learning and the Prototypical Network approach to this problem.
|
| 27 |
+
|
| 28 |
+
# 2.1 FEW-SHOT LEARNING
|
| 29 |
+
|
| 30 |
+
Recent progress on few-shot learning has been made possible by following an episodic paradigm. Consider a situation where we have a large labeled dataset for a set of classes $\mathcal { C } _ { \mathrm { t r a i n } }$ . However, after training on examples from $\mathcal { C } _ { \mathrm { t r a i n } }$ , our ultimate goal is to produce classifiers for a disjoint set of new classes $\mathcal { C } _ { \mathrm { t e s t } }$ , for which only a few labeled examples will be available. The idea behind the episodic paradigm is to simulate the types of few-shot problems that will be encountered at test, taking advantage of the large quantities of available labeled data for classes $\mathcal { C } _ { \mathrm { t r a i n } }$ .
|
| 31 |
+
|
| 32 |
+
Specifically, models are trained on $K$ -shot, $N$ -way episodes constructed by first sampling a small subset of $N$ classes from $\mathcal { C } _ { \mathrm { t r a i n } }$ and then generating: 1) a training (support) set ${ \boldsymbol { s } } \ =$ $\left\{ ( \pmb { x } _ { 1 } , y _ { 1 } ) , ( \pmb { x } _ { 2 } , y _ { 2 } ) , \dots , ( \pmb { x } _ { N \times K } , y _ { N \times K } ) \right\}$ containing $K$ examples from each of the $N$ classes and 2) a test (query) set $\mathcal { Q } = \{ ( \pmb { x } _ { 1 } ^ { * } , y _ { 1 } ^ { * } ) , ( \pmb { x } _ { 2 } ^ { * } , y _ { 2 } ^ { * } ) , . . . , ( \pmb { x } _ { T } ^ { * } , y _ { T } ^ { * } ) \}$ of different examples from the same $N$ classes. Each $\pmb { x } _ { i } \in \mathbb { R } ^ { D }$ is an input vector of dimension $D$ and $y _ { i } \in \{ 1 , 2 , \ldots , N \}$ is a class label (similarly for $\boldsymbol { \mathscr { x } } _ { i } ^ { * }$ and $y _ { i } ^ { * }$ ). Training on such episodes is done by feeding the support set $s$ to the model and updating its parameters to minimize the loss of its predictions for the examples in the query set $\mathcal { Q }$ .
|
| 33 |
+
|
| 34 |
+
One way to think of this approach is that our model effectively trains to be a good learning algorithm. Indeed, much like a learning algorithm, the model must take in a set of labeled examples and produce a predictor that can be applied to new examples. Moreover, training directly encourages the classifier produced by the model to have good generalization on the new examples of the query set. Due to this analogy, training under this paradigm is often referred to as learning to learn or meta-learning.
|
| 35 |
+
|
| 36 |
+
On the other hand, referring to the content of episodes as training and test sets and to the process of learning on these episodes as meta-learning or meta-training (as is sometimes done in the literature) can be confusing. So for the sake of clarity, we will refer to the content of episodes as support and query sets, and to the process of iterating over the training episodes simply as training.
|
| 37 |
+
|
| 38 |
+
# 2.2 PROTOTYPICAL NETWORKS
|
| 39 |
+
|
| 40 |
+
Prototypical Network (Snell et al., 2017) is a few-shot learning model that has the virtue of being simple and yet obtaining state-of-the-art performance. At a high-level, it uses the support set $s$ to extract a prototype vector from each class, and classifies the inputs in the query set based on their distance to the prototype of each class.
|
| 41 |
+
|
| 42 |
+
More precisely, Prototypical Networks learn an embedding function $h ( { \pmb x } )$ , parameterized as a neural network, that maps examples into a space where examples from the same class are close and those from different classes are far. All parameters of Prototypical Networks lie in the embedding function.
|
| 43 |
+
|
| 44 |
+
To compute the prototype $\pmb { p } _ { c }$ of each class $c$ , a per-class average of the embedded examples is performed:
|
| 45 |
+
|
| 46 |
+
$$
|
| 47 |
+
p _ { c } = \frac { \sum _ { i } h ( \pmb { x } _ { i } ) z _ { i , c } } { \sum _ { i } z _ { i , c } } , \mathrm { w h e r e } z _ { i , c } = \mathbb { 1 } [ y _ { i } = c ] .
|
| 48 |
+
$$
|
| 49 |
+
|
| 50 |
+
These prototypes define a predictor for the class of any new (query) example $\mathbf { \boldsymbol { x } } ^ { * }$ , which assigns a probability over any class $c$ based on the distances between $\pmb { x } ^ { * }$ and each prototype, as follows:
|
| 51 |
+
|
| 52 |
+
$$
|
| 53 |
+
p ( c | \pmb { x } ^ { * } , \{ p _ { c } \} ) = \frac { \exp ( - | | h ( \pmb { x } ^ { * } ) - \pmb { p } _ { c } | | _ { 2 } ^ { 2 } ) } { \sum _ { c ^ { \prime } } \exp ( - | | h ( \pmb { x } ^ { * } ) - \pmb { p } _ { c ^ { \prime } } | | _ { 2 } ^ { 2 } ) } .
|
| 54 |
+
$$
|
| 55 |
+
|
| 56 |
+
The loss function used to update Prototypical Networks for a given training episode is then simply the average negative log-probability of the correct class assignments, for all query examples:
|
| 57 |
+
|
| 58 |
+
$$
|
| 59 |
+
- \frac { 1 } { T } \sum _ { i } \log p ( y _ { i } ^ { * } | \pmb { x } _ { i } ^ { * } , \{ \pmb { p } _ { c } \} ) .
|
| 60 |
+
$$
|
| 61 |
+
|
| 62 |
+
Training proceeds by minimizing the average loss, iterating over training episodes and performing a gradient descent update for each.
|
| 63 |
+
|
| 64 |
+
Generalization performance is measured on test set episodes, which contain images from classes in $\mathcal { C } _ { \mathrm { t e s t } }$ instead of $\mathcal { C } _ { \mathrm { t r a i n } }$ . For each test episode, we use the predictor produced by the Prototypical Network for the provided support set $s$ to classify each of query input $\pmb { x } ^ { * }$ into the most likely class $\hat { y } = \operatorname { a r g m a x } _ { c } p ( \bar { c } | \pmb { x } ^ { * } , \{ \pmb { p } _ { c } \} )$ .
|
| 65 |
+
|
| 66 |
+
# 3 SEMI-SUPERVISED FEW-SHOT LEARNING
|
| 67 |
+
|
| 68 |
+
We now define the semi-supervised setting considered in this work for few-shot learning.
|
| 69 |
+
|
| 70 |
+
The training set is denoted as a tuple of labeled and unlabeled examples: $( S , R )$ . The labeled portion is the usual support set $s$ of the few-shot learning literature, containing a list of tuples of inputs and targets. In addition to classic few-shot learning, we introduce an unlabeled set $\mathcal { R }$ containing only inputs: $\mathcal { R } = \{ \tilde { { \pmb { x } } } _ { 1 } , \tilde { { \pmb { x } } } _ { 2 } , \ldots , \tilde { { \pmb { x } } } _ { M } \}$ . As in the purely supervised setting, our models are trained to perform well when predicting the labels for the examples in the episode’s query set $\mathcal { Q }$ . Figure 2 shows a visualization of training and test episodes.
|
| 71 |
+
|
| 72 |
+

|
| 73 |
+
Figure 2: Example of the semi-supervised few-shot learning setup. Training involves iterating through training episodes, consisting of a support set $s$ , an unlabeled set $\mathcal { R }$ , and a query set $\mathcal { Q }$ . The goal is to use the labeled items (shown with their numeric class label) in $s$ and the unlabeled items in $\mathcal { R }$ within each episode to generalize to good performance on the corresponding query set. The unlabeled items in $\mathcal { R }$ may either be pertinent to the classes we are considering (shown above with green plus signs) or they may be distractor items which belong to a class that is not relevant to the current episode (shown with red minus signs). However note that the model does not actually have ground truth information as to whether each unlabeled example is a distractor or not; the plus/minus signs are shown only for illustrative purposes. At test time, we are given new episodes consisting of novel classes not seen during training that we use to evaluate the meta-learning method.
|
| 74 |
+
|
| 75 |
+
# 3.1 SEMI-SUPERVISED PROTOTYPICAL NETWORKS
|
| 76 |
+
|
| 77 |
+
In their original formulation, Prototypical Networks do not specify a way to leverage the unlabeled set $\mathcal { R }$ . In what follows, we now propose various extensions that start from the basic definition of prototypes $\pmb { p _ { c } }$ and provide a procedure for producing refined prototypes $\tilde { p } _ { c }$ using the unlabeled examples in $\mathcal { R }$ .
|
| 78 |
+
|
| 79 |
+
After the refined prototypes are obtained, each of these models is trained with the same loss function for ordinary Prototypical Networks of Equation 3, but replacing ${ \pmb p } _ { c }$ with $\tilde { p } _ { c }$ . That is, each query example is classified into one of the $N$ classes based on the proximity of its embedded position with the corresponding refined prototypes, and the average negative logprobability of the correct classification is used for training.
|
| 80 |
+
|
| 81 |
+

|
| 82 |
+
Figure 3: Left: The prototypes are initialized based on the mean location of the examples of the corresponding class, as in ordinary Prototypical Networks. Support, unlabeled, and query examples have solid, dashed, and white colored borders respectively. Right: The refined prototypes obtained by incorporating the unlabeled examples, which classifies all query examples correctly.
|
| 83 |
+
|
| 84 |
+
# 3.1.1 PROTOTYPICAL NETWORKS WITH SOFT $k$ -MEANS
|
| 85 |
+
|
| 86 |
+
We first consider a simple way of leveraging unlabeled examples for refining prototypes, by taking inspiration from semi-supervised clustering. Viewing each prototype as a cluster center, the refinement process could attempt to adjust the cluster locations to better fit the examples in both the support and unlabeled sets. Under this view, cluster assignments of the labeled examples in the support set are considered known and fixed to each example’s label. The refinement process must instead estimate the cluster assignments of the unlabeled examples and adjust the cluster locations (the prototypes) accordingly.
|
| 87 |
+
|
| 88 |
+
One natural choice would be to borrow from the inference performed by soft $k$ -means. We prefer this version of $k$ -means over hard assignments since hard assignments would make the inference non-differentiable. We start with the regular Prototypical Network’s prototypes $\pmb { p _ { c } }$ (as specified in Equation 1) as the cluster locations. Then, the unlabeled examples get a partial assignment $( \tilde { z } _ { j , c } )$ to each cluster based on their Euclidean distance to the cluster locations. Finally, refined prototypes are obtained by incorporating these unlabeled examples.
|
| 89 |
+
|
| 90 |
+
This process can be summarized as follows:
|
| 91 |
+
|
| 92 |
+
$$
|
| 93 |
+
\tilde { p } _ { c } = \frac { \sum _ { i } h ( { \bf x } _ { i } ) z _ { i , c } + \sum _ { j } h ( \tilde { x } _ { j } ) \tilde { z } _ { j , c } } { \sum _ { i } z _ { i , c } + \sum _ { j } \tilde { z } _ { j , c } } , \mathrm { w h e r e } z _ { j , c } = \frac { \exp \left( - | | h ( \tilde { x } _ { j } ) - p _ { c } | | _ { 2 } ^ { 2 } \right) } { \sum _ { c ^ { \prime } } \exp \left( - | | h ( \tilde { x } _ { j } ) - p _ { c ^ { \prime } } | | _ { 2 } ^ { 2 } \right) }
|
| 94 |
+
$$
|
| 95 |
+
|
| 96 |
+
Predictions of each query input’s class is then modeled as in Equation 2, but using the refined prototypes $\tilde { p } _ { c }$ .
|
| 97 |
+
|
| 98 |
+
We could perform several iterations of refinement, as is usual in $k$ -means. However, we have experimented with various number of iterations and found results to not improve beyond a single refinement step.
|
| 99 |
+
|
| 100 |
+
# 3.1.2 PROTOTYPICAL NETWORKS WITH SOFT $k$ -MEANS WITH A DISTRACTOR CLUSTER
|
| 101 |
+
|
| 102 |
+
The soft $k$ -means approach described above implicitly assumes that each unlabeled example belongs to either one of the $N$ classes in the episode. However, it would be much more general to not make that assumption and have a model robust to the existence of examples from other classes, which we refer to as distractor classes. For example, such a situation would arise if we wanted to distinguish between pictures of unicycles and scooters, and decided to add an unlabeled set by downloading images from the web. It then would not be realistic to assume that all these images are of unicycles or scooters. Even with a focused search, some may be from similar classes, such as bicycle.
|
| 103 |
+
|
| 104 |
+
Since soft $k$ -means distributes its soft assignments across all classes, distractor items could be harmful and interfere with the refinement process, as prototypes would be adjusted to also partially account for these distractors. A simple way to address this is to add an additional cluster whose purpose is to capture the distractors, thus preventing them from polluting the clusters of the classes of interest:
|
| 105 |
+
|
| 106 |
+
$$
|
| 107 |
+
\pmb { p _ { c } } = \left\{ \begin{array} { l l } { \frac { \sum _ { i } h ( \pmb { x _ { i } } ) z _ { i , c } } { \sum _ { i } z _ { i , c } } } & { \quad \mathrm { f o r } c = 1 . . . N } \\ { \mathbf { 0 } } & { \quad \mathrm { f o r } c = N + 1 } \end{array} \right.
|
| 108 |
+
$$
|
| 109 |
+
|
| 110 |
+
Here we take the simplifying assumption that the distractor cluster has a prototype centered at the origin. We also consider introducing length-scales $r _ { c }$ to represent variations in the within-cluster distances, specifically for the distractor cluster:
|
| 111 |
+
|
| 112 |
+
$$
|
| 113 |
+
\tilde { z } _ { j , c } = \frac { \exp \left( - \frac { 1 } { r _ { c } ^ { 2 } } | | \tilde { x } _ { j } - p _ { c } | | _ { 2 } ^ { 2 } - A ( r _ { c } ) \right) } { \sum _ { c ^ { \prime } } \exp \left( - \frac { 1 } { r _ { c } ^ { 2 } } | | \tilde { x } _ { j } - p _ { c ^ { \prime } } | | _ { 2 } ^ { 2 } - A ( r _ { c ^ { \prime } } ) \right) } , \mathrm { ~ w h e r e ~ } A ( r ) = \frac { 1 } { 2 } \log ( 2 \pi ) + \log ( r )
|
| 114 |
+
$$
|
| 115 |
+
|
| 116 |
+
For simplicity, we set $r _ { 1 \dots N }$ to 1 in our experiments, and only learn the length-scale of the distractor cluster rN+1.
|
| 117 |
+
|
| 118 |
+
# 3.1.3 PROTOTYPICAL NETWORKS WITH SOFT $k$ -MEANS AND MASKING
|
| 119 |
+
|
| 120 |
+
Modeling distractor unlabeled examples with a single cluster is likely too simplistic. Indeed, it is inconsistent with our assumption that each cluster corresponds to one class, since distractor examples may very well cover more than a single natural object category. Continuing with our unicycles and bicycles example, our web search for unlabeled images could accidentally include not only bicycles, but other related objects such as tricycles or cars. This was also reflected in our experiments, where we constructed the episode generating process so that it would sample distractor examples from multiple classes.
|
| 121 |
+
|
| 122 |
+
To address this problem, we propose an improved variant: instead of capturing distractors with a high-variance catch-all cluster, we model distractors as examples that are not within some area of any of the legitimate class prototypes. This is done by incorporating a soft-masking mechanism on the contribution of unlabeled examples. At a high level, we want unlabeled examples that are closer to a prototype to be masked less than those that are farther.
|
| 123 |
+
|
| 124 |
+
More specifically, we modify the soft $k$ -means refinement as follows. We start by computing normalized distances $\tilde { d } _ { j , c }$ between examples $\tilde { \mathbfit { x } } _ { j }$ and prototypes $\pmb { p _ { c } }$ :
|
| 125 |
+
|
| 126 |
+
$$
|
| 127 |
+
\tilde { d } _ { j , c } = \frac { d _ { j , c } } { \frac { 1 } { M } \sum _ { j } d _ { j , c } } , \ \mathrm { w h e r e } \ d _ { j , c } = | | h ( \tilde { \bf x } _ { j } ) - { \bf p } _ { c } | | _ { 2 } ^ { 2 }
|
| 128 |
+
$$
|
| 129 |
+
|
| 130 |
+
Then, soft thresholds $\beta _ { c }$ and slopes $\gamma _ { c }$ are predicted for each prototype, by feeding to a small neural network various statistics of the normalized distances for the prototype:
|
| 131 |
+
|
| 132 |
+
$$
|
| 133 |
+
[ \beta _ { c } , \gamma _ { c } ] = \mathrm { M L P } \left( \left[ \operatorname* { m i n } _ { j } ( \tilde { d } _ { j , c } ) , \operatorname* { m a x } _ { j } ( \tilde { d } _ { j , c } ) , \operatorname { v a r } _ { j } ( \tilde { d } _ { j , c } ) , \operatorname { s k e w } ( \tilde { d } _ { j , c } ) , \operatorname { k u r t } _ { j } ( \tilde { d } _ { j , c } ) \right] \right)
|
| 134 |
+
$$
|
| 135 |
+
|
| 136 |
+
This allows each threshold to use information on the amount of intra-cluster variation to determine how aggressively it should cut out unlabeled examples.
|
| 137 |
+
|
| 138 |
+
Then, soft masks $m _ { j , c }$ for the contribution of each example to each prototype are computed, by comparing to the threshold the normalized distances, as follows:
|
| 139 |
+
|
| 140 |
+
$$
|
| 141 |
+
\tilde { p } _ { c } = \frac { \sum _ { i } h ( { \bf x } _ { i } ) z _ { i , c } + \sum _ { j } h ( \tilde { x } _ { j } ) \tilde { z } _ { j , c } m _ { j , c } } { \sum _ { i } z _ { i , c } + \sum _ { j } \tilde { z } _ { j , c } m _ { j , c } } , \mathrm { ~ w h e r e ~ } m _ { j , c } = \sigma \left( - \gamma _ { c } \left( \tilde { d } _ { j , c } - \beta _ { c } \right) \right)
|
| 142 |
+
$$
|
| 143 |
+
|
| 144 |
+
where $\sigma ( \cdot )$ is the sigmoid function.
|
| 145 |
+
|
| 146 |
+
When training with this refinement process, the model can now use its MLP in Equation 8 to learn to include or ignore entirely certain unlabeled examples. The use of soft masks makes this process entirely differentiable2. Finally, much like for regular soft $k$ -means (with or without a distractor cluster), while we could recursively repeat the refinement for multiple steps, we found a single step to perform well enough.
|
| 147 |
+
|
| 148 |
+
# 4 RELATED WORK
|
| 149 |
+
|
| 150 |
+
We summarize here the most relevant work from the literature on few-shot learning, semi-supervised learning and clustering.
|
| 151 |
+
|
| 152 |
+
The best performing methods for few-shot learning use the episodic training framework prescribed by meta-learning. The approach within which our work falls is that of metric learning methods. Previous work in metric-learning for few-shot-classification includes Deep Siamese Networks (Koch et al., 2015), Matching Networks (Vinyals et al., 2016), and Prototypical Networks (Snell et al., 2017), which is the model we extend to the semi-supervised setting in our work. The general idea here is to learn an embedding function that embeds examples belonging to the same class close together while keeping embeddings from separate classes far apart. Distances between embeddings of items from the support set and query set are then used as a notion of similarity to do classification. Lastly, closely related to our work with regard to extending the few-shot learning setting, Bachman et al. (2017) employ Matching Networks in an active learning framework where the model has a choice of which unlabeled item to add to the support set over a certain number of time steps before classifying the query set. Unlike our setting, their meta-learning agent can acquire ground-truth labels from the unlabeled set, and they do not use distractor examples.
|
| 153 |
+
|
| 154 |
+
Other meta-learning approaches to few-shot learning include learning how to use the support set to update a learner model so as to generalize to the query set. Recent work has involved learning either the weight initialization and/or update step that is used by a learner neural network (Ravi & Larochelle, 2017; Finn et al., 2017). Another approach is to train a generic neural architecture such as a memory-augmented recurrent network (Santoro et al., 2016) or a temporal convolutional network (Mishra et al., 2017) to sequentially process the support set and perform accurate predictions of the labels of the query set examples. These other methods are also competitive for few-shot learning, but we chose to extend Prototypical Networks in this work for its simplicity and efficiency.
|
| 155 |
+
|
| 156 |
+
As for the literature on semi-supervised learning, while it is quite vast (Zhu, 2005; Chapelle et al., 2010), the most relevant category to our work is related to self-training (Yarowsky, 1995; Rosenberg et al., 2005). Here, a classifier is first trained on the initial training set. The classifier is then used to classify unlabeled items, and the most confidently predicted unlabeled items are added to the training set with the prediction of the classifier as the assumed label. This is similar to our soft $k$ -Means extension to Prototypical Networks. Indeed, since the soft assignments (Equation 4) match the regular Prototypical Network’s classifier output for new inputs (Equation 2), then the refinement can be thought of re-feeding to a Prototypical Network a new support set augmented with (soft) self-labels from the unlabeled set.
|
| 157 |
+
|
| 158 |
+
Our algorithm is also related to transductive learning (Vapnik, 1998; Joachims, 1999; Fu et al., 2015), where the base classifier gets refined by seeing the unlabeled examples. In practice, one could use our method in a transductive setting where the unlabeled set is the same as the query set; however, here to avoid our model memorizing labels of the unlabeled set during the meta-learning procedure, we split out a separate unlabeled set that is different from the query set.
|
| 159 |
+
|
| 160 |
+
In addition to the original $k$ -Means method (Lloyd, 1982), the most related work to our setup involving clustering algorithms considers applying $k$ -Means in the presence of outliers (Hautamäki et al., 2005; Chawla & Gionis, 2013; Gupta et al., 2017). The goal here is to correctly discover and ignore the outliers so that they do not wrongly shift the cluster locations to form a bad partition of the true data. This objective is also important in our setup as not ignoring outliers (or distractors) will wrongly shift the prototypes and negatively influence classification performance.
|
| 161 |
+
|
| 162 |
+
Our contribution to the semi-supervised learning and clustering literature is to go beyond the classical setting of training and evaluating within a single dataset, and consider the setting where we must learn to transfer from a set of training classes $\mathcal { C } _ { \mathrm { t r a i n } }$ to a new set of test classes $\mathcal { C } _ { \mathrm { t e s t } }$ .
|
| 163 |
+
|
| 164 |
+
# 5 EXPERIMENTS
|
| 165 |
+
|
| 166 |
+
# 5.1 DATASETS
|
| 167 |
+
|
| 168 |
+
We evaluate the performance of our model on three datasets: two benchmark few-shot classification datasets and a novel large-scale dataset that we hope will be useful for future few-shot learning work.
|
| 169 |
+
|
| 170 |
+
Omniglot (Lake et al., 2011) is a dataset of 1,623 handwritten characters from 50 alphabets. Each character was drawn by 20 human subjects. We follow the few-shot setting proposed by Vinyals et al. (2016), in which the images are resized to $2 8 \times 2 8$ pixels and rotations in multiples of $9 0 ^ { \circ }$ are applied, yielding 6,492 classes in total. These are split into 4,112 training classes, 688 validation classes, and 1,692 testing classes.
|
| 171 |
+
|
| 172 |
+
miniImageNet (Vinyals et al., 2016) is a modified version of the ILSVRC-12 dataset (Russakovsky et al., 2015), in which 600 images for each of 100 classes were randomly chosen to be part of the dataset. We rely on the class split used by Ravi & Larochelle (2017). These splits use 64 classes for training, 16 for validation, and 20 for test. All images are of size $8 4 \times 8 4$ pixels.
|
| 173 |
+
|
| 174 |
+
tieredImageNet is our proposed dataset for few-shot classification. Like miniImagenet, it is a subset of ILSVRC-12. However, tieredImageNet represents a larger subset of ILSVRC-12 (608 classes rather than 100 for miniImageNet). Analogous to Omniglot, in which characters are grouped into alphabets, tieredImageNet groups classes into broader categories corresponding to higher-level nodes in the ImageNet (Deng et al., 2009) hierarchy. There are 34 categories in total, with each category containing between 10 and 30 classes. These are split into 20 training, 6 validation and 8 testing categories (details of the dataset can be found in the supplementary material). This ensures that all of the training classes are sufficiently distinct from the testing classes, unlike miniImageNet and other alternatives such as randImageNet proposed by Vinyals et al. (2016). For example, “pipe organ” is a training class and “electric guitar” is a test class in the Ravi & Larochelle (2017) split of miniImagenet, even though they are both musical instruments. This scenario would not occur in tieredImageNet since “musical instrument” is a high-level category and as such is not split between training and test classes. This represents a more realistic few-shot learning scenario since in general we cannot assume that test classes will be similar to those seen in training. Additionally, the tiered structure of tieredImageNet may be useful for few-shot learning approaches that can take advantage of hierarchical relationships between classes. We leave such interesting extensions for future work.
|
| 175 |
+
|
| 176 |
+
Table 1: Omniglot 1-shot classification results. In this table as well as those below “w/ D” denotes “with distractors”, where the unlabeled images contain irrelevant classes.
|
| 177 |
+
|
| 178 |
+
<table><tr><td rowspan=1 colspan=1>Models</td><td rowspan=1 colspan=1>Acc.</td><td rowspan=1 colspan=1>Acc. w/ D</td></tr><tr><td rowspan=1 colspan=1>Supervised</td><td rowspan=1 colspan=1>94.62 ± 0.09</td><td rowspan=1 colspan=1>94.62 ± 0.09</td></tr><tr><td rowspan=1 colspan=1>Semi-Supervised Inference</td><td rowspan=1 colspan=1>97.45 ± 0.05</td><td rowspan=1 colspan=1>95.08 ± 0.09</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means</td><td rowspan=1 colspan=1>97.25 ± 0.10</td><td rowspan=1 colspan=1>95.01 ± 0.09</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means+Cluster</td><td rowspan=1 colspan=1>97.68 ± 0.07</td><td rowspan=1 colspan=1>97.17 ± 0.04</td></tr><tr><td rowspan=1 colspan=1>Masked Soft k-Means</td><td rowspan=1 colspan=1>97.52 ± 0.07</td><td rowspan=1 colspan=1>97.30 ± 0.08</td></tr></table>
|
| 179 |
+
|
| 180 |
+
# 5.2 ADAPTING THE DATASETS FOR SEMI-SUPERVISED LEARNING
|
| 181 |
+
|
| 182 |
+
For each dataset, we first create an additional split to separate the images of each class into disjoint labeled and unlabeled sets. For Omniglot and tieredImageNet we sample $10 \%$ of the images of each class to form the labeled split. The remaining $90 \%$ can only be used in the unlabeled portion of episodes. For miniImageNet we use $40 \%$ of the data for the labeled split and the remaining $60 \%$ for the unlabeled, since we noticed that $10 \%$ was too small to achieve reasonable performance and avoid overfitting. We report the average classification scores over 10 random splits of labeled and unlabeled portions of the training set, with uncertainty computed in standard error (standard deviation divided by the square root of the total number of splits).
|
| 183 |
+
|
| 184 |
+
We would like to emphasize that due to this labeled/unlabeled split, we are using strictly less label information than in the previously-published work on these datasets. Because of this, we do not expect our results to match the published numbers, which should instead be interpreted as an upperbound for the performance of the semi-supervised models defined in this work.
|
| 185 |
+
|
| 186 |
+
Episode construction then is performed as follows. For a given dataset, we create a training episode by first sampling $N$ classes uniformly at random from the set of training classes $\mathcal { C } _ { \mathrm { t r a i n } }$ . We then sample $K$ images from the labeled split of each of these classes to form the support set, and $M$ images from the unlabeled split of each of these classes to form the unlabeled set. Optionally, when including distractors, we additionally sample $H$ other classes from the set of training classes and $M$ images from the unlabeled split of each to act as the distractors. These distractor images are added to the unlabeled set along with the unlabeled images of the $N$ classes of interest (for a total of $M N + M H$ unlabeled images). The query portion of the episode is comprised of a fixed number of images from the labeled split of each of the $N$ chosen classes. Test episodes are created analogously, but with the $N$ classes (and optionally the $H$ distractor classes) sampled from $\mathcal { C } _ { \mathrm { t e s t } }$ . In the experiments reported here we used $H = N = 5$ , i.e. 5 classes for both the labeled classes and the distractor classes. We used $M = 5$ for training and $M = 2 0$ for testing in most cases, thus measuring the ability of the models to generalize to a larger unlabeled set size. Details of the dataset splits, including the specific classes assigned to train/validation/test sets, can be found in Appendices A and B.
|
| 187 |
+
|
| 188 |
+
In each dataset we compare our three semi-supervised models with two baselines. The first baseline, referred to as “Supervised” in our tables, is an ordinary Prototypical Network that is trained in a purely supervised way on the labeled split of each dataset. The second baseline, referred to as “Semi-Supervised Inference”, uses the embedding function learned by this supervised Prototypical Network, but performs semi-supervised refinement of the prototypes at test time using a step of Soft $k$ -Means refinement. This is to be contrasted with our semi-supervised models that perform this refinement both at training time and at test time, therefore learning a different embedding function. We evaluate each model in two settings: one where all unlabeled examples belong to the classes of interest, and a more challenging one that includes distractors. Details of the model hyperparameters can be found in Appendix D and our online repository.3
|
| 189 |
+
|
| 190 |
+
# 5.3 RESULTS
|
| 191 |
+
|
| 192 |
+
Results for Omniglot, miniImageNet and tieredImageNet are given in Tables 1, 2 and 5, respectively, while Figure 4 shows the performance of our models on tieredImageNet (our largest dataset) using different values for $M$ (number of items in the unlabeled set per class). Additional results comparing the ProtoNet model to various baselines on these datasets, and analysis of the performance of the Masked Soft $k$ -Means model can be found in Appendix C.
|
| 193 |
+
|
| 194 |
+
Table 2: miniImageNet 1/5-shot classification results.
|
| 195 |
+
|
| 196 |
+
<table><tr><td rowspan=1 colspan=1>Models</td><td rowspan=1 colspan=1>1-shot Acc.</td><td rowspan=1 colspan=1>5-shot Acc.</td><td rowspan=1 colspan=1>1-shot Acc w/ D</td><td rowspan=1 colspan=1> 5-shot Acc. w/ D</td></tr><tr><td rowspan=1 colspan=1>Supervised</td><td rowspan=1 colspan=1>43.61 ± 0.27</td><td rowspan=1 colspan=1>59.08 ± 0.22</td><td rowspan=1 colspan=1>43.61 ± 0.27</td><td rowspan=1 colspan=1>59.08 ± 0.22</td></tr><tr><td rowspan=1 colspan=1>Semi-Supervised Inference</td><td rowspan=1 colspan=1>48.98 ± 0.34</td><td rowspan=1 colspan=1>63.77 ± 0.20</td><td rowspan=1 colspan=1>47.42 ± 0.33</td><td rowspan=1 colspan=1>62.62 ± 0.24</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means</td><td rowspan=1 colspan=1>50.09 ± 0.45</td><td rowspan=1 colspan=1>64.59 ± 0.28</td><td rowspan=1 colspan=1>48.70 ± 0.32</td><td rowspan=1 colspan=1>63.55 ± 0.28</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means+Cluster</td><td rowspan=1 colspan=1>49.03 ± 0.24</td><td rowspan=1 colspan=1>63.08 ± 0.18</td><td rowspan=1 colspan=1>48.86± 0.32</td><td rowspan=1 colspan=1>61.27 ± 0.24</td></tr><tr><td rowspan=1 colspan=1>Masked Soft k-Means</td><td rowspan=1 colspan=1>50.41 ± 0.31</td><td rowspan=1 colspan=1>64.39 ± 0.24</td><td rowspan=1 colspan=1>49.04 ± 0.31</td><td rowspan=1 colspan=1>62.96 ± 0.14</td></tr></table>
|
| 197 |
+
|
| 198 |
+
Table 3: tieredImageNet 1/5-shot classification results.
|
| 199 |
+
|
| 200 |
+
<table><tr><td rowspan=1 colspan=1>Models</td><td rowspan=1 colspan=1>1-shot Acc.</td><td rowspan=1 colspan=1> 5-shot Acc.</td><td rowspan=1 colspan=1>1-shot Acc. w/ D</td><td rowspan=1 colspan=1> 5-shot Acc. w/ D</td></tr><tr><td rowspan=1 colspan=1>Supervised</td><td rowspan=1 colspan=1>46.52 ± 0.52</td><td rowspan=1 colspan=1>66.15 ± 0.22</td><td rowspan=1 colspan=1>46.52 ± 0.52</td><td rowspan=1 colspan=1>66.15 ± 0.22</td></tr><tr><td rowspan=1 colspan=1>Semi-Supervised Inference</td><td rowspan=1 colspan=1>50.74 ± 0.75</td><td rowspan=1 colspan=1>69.37 ± 0.26</td><td rowspan=1 colspan=1>48.67 ± 0.60</td><td rowspan=1 colspan=1>67.46 ± 0.24</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means</td><td rowspan=1 colspan=1>51.52 ± 0.36</td><td rowspan=1 colspan=1>70.25 ± 0.31</td><td rowspan=1 colspan=1>49.88 ± 0.52</td><td rowspan=1 colspan=1>68.32 ± 0.22</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means+Cluster</td><td rowspan=1 colspan=1>51.85 ± 0.25</td><td rowspan=1 colspan=1>69.42 ± 0.17</td><td rowspan=1 colspan=1>51.36 ± 0.31</td><td rowspan=1 colspan=1>67.56 ± 0.10</td></tr><tr><td rowspan=1 colspan=1>Masked Soft k-Means</td><td rowspan=1 colspan=1>52.39 ± 0.44</td><td rowspan=1 colspan=1>69.88 ± 0.20</td><td rowspan=1 colspan=1>51.38 ± 0.38</td><td rowspan=1 colspan=1>69.08 ± 0.25</td></tr></table>
|
| 201 |
+
|
| 202 |
+
Across all three benchmarks, at least one of our proposed models outperforms the baselines, demonstrating the effectiveness of our semi-supervised meta-learning procedure. In the nondistractor settings, all three proposed models outperform the baselines in almost all the experiments, without a clear winner between the three models across the datasets and shot numbers. In the scenario where training and testing includes distractors, Masked Soft $k$ -Means shows the most robust performance across all three datasets, attaining the best results in each case but one. In fact this model reaches performance that is close to the upper bound based on the results without distractors.
|
| 203 |
+
|
| 204 |
+
From Figure 4, we observe clear improvements in test accuracy when the number of items in the unlabeled set per class grows from 0 to 25. These models were trained with $M = 5$ and thus are showing an ability to extrapolate in generalization. This confirms that, through meta-training, the models learn to acquire a better representation that is improved by semi-supervised refinement.
|
| 205 |
+
|
| 206 |
+
# 6 CONCLUSION
|
| 207 |
+
|
| 208 |
+
In this work, we propose a novel semi-supervised few-shot learning paradigm, where an unlabeled set is added to each episode. We also extend the setup to more realistic situations where the unlabeled set has novel classes distinct from the labeled classes. To address the problem that current fewshot classification datasets are too small for a labeled vs. unlabeled split and also lack hierarchical levels of labels, we introduce a new dataset, tieredImageNet. We propose several novel extensions of Prototypical Networks, and they show consistent improvements under semi-supervised settings compared to our baselines. As future work, we are working on incorporating fast weights (Ba et al., 2016; Finn et al., 2017) into our framework so that examples can have different embedding representations given the contents in the episode.
|
| 209 |
+
|
| 210 |
+
Acknowledgement Supported by grants from NSERC, Samsung, and the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DoI/IBC) contract number D16PC00003. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DoI/IBC, or the U.S. Government.
|
| 211 |
+
|
| 212 |
+

|
| 213 |
+
Figure 4: Model Performance on tieredImageNet with different numbers of unlabeled items during test time.
|
| 214 |
+
|
| 215 |
+
# REFERENCES
|
| 216 |
+
|
| 217 |
+
Jimmy Ba, Geoffrey E. Hinton, Volodymyr Mnih, Joel Z. Leibo, and Catalin Ionescu. Using fast weights to attend to the recent past. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pp. 4331–4339, 2016.
|
| 218 |
+
|
| 219 |
+
Philip Bachman, Alessandro Sordoni, and Adam Trischler. Learning algorithms for active learning. 2017.
|
| 220 |
+
|
| 221 |
+
Olivier Chapelle, Bernhard Schölkopf, and Alexander Zien. Semi-Supervised Learning. The MIT Press, 1st edition, 2010. ISBN 0262514125, 9780262514125.
|
| 222 |
+
|
| 223 |
+
Sanjay Chawla and Aristides Gionis. k-means–: A unified approach to clustering and outlier detection. In Proceedings of the 2013 SIAM International Conference on Data Mining, pp. 189– 197. SIAM, 2013.
|
| 224 |
+
|
| 225 |
+
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pp. 248–255. IEEE, 2009.
|
| 226 |
+
|
| 227 |
+
Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In 34th International Conference on Machine Learning, 2017.
|
| 228 |
+
|
| 229 |
+
Yanwei Fu, Timothy M. Hospedales, Tao Xiang, and Shaogang Gong. Transductive multi-view zero-shot learning. IEEE Trans. Pattern Anal. Mach. Intell., 37(11):2332–2345, 2015.
|
| 230 |
+
|
| 231 |
+
Shalmoli Gupta, Ravi Kumar, Kefu Lu, Benjamin Moseley, and Sergei Vassilvitskii. Local search methods for k-means with outliers. Proceedings of the VLDB Endowment, 10(7):757–768, 2017.
|
| 232 |
+
|
| 233 |
+
Ville Hautamäki, Svetlana Cherednichenko, Ismo Kärkkäinen, Tomi Kinnunen, and Pasi Fränti. Improving k-means by outlier removal. In Scandinavian Conference on Image Analysis, pp. 978– 987. Springer, 2005.
|
| 234 |
+
|
| 235 |
+
Sepp Hochreiter, A Steven Younger, and Peter R Conwell. Learning to learn using gradient descent. In International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001.
|
| 236 |
+
|
| 237 |
+
Thorsten Joachims. Transductive inference for text classification using support vector machines. In Proceedings of the Sixteenth International Conference on Machine Learning, 1999.
|
| 238 |
+
|
| 239 |
+
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
|
| 240 |
+
|
| 241 |
+
Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. Siamese neural networks for one-shot image recognition. In ICML Deep Learning Workshop, volume 2, 2015.
|
| 242 |
+
|
| 243 |
+
Brenden M. Lake, Ruslan Salakhutdinov, Jason Gross, and Joshua B. Tenenbaum. One shot learning of simple visual concepts. In Proceedings of the 33th Annual Meeting of the Cognitive Science Society, CogSci 2011, Boston, Massachusetts, USA, July 20-23, 2011, 2011.
|
| 244 |
+
|
| 245 |
+
Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory, 28(2): 129–137, 1982.
|
| 246 |
+
|
| 247 |
+
Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. Meta-learning with temporal convolutions. CoRR, abs/1707.03141, 2017.
|
| 248 |
+
|
| 249 |
+
Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In 5th International Conference on Learning Representations, 2017.
|
| 250 |
+
|
| 251 |
+
Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. 2005.
|
| 252 |
+
|
| 253 |
+
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.
|
| 254 |
+
|
| 255 |
+
Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy P. Lillicrap. Oneshot learning with memory-augmented neural networks. In 33rd International Conference on Machine Learning, 2016.
|
| 256 |
+
|
| 257 |
+
Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems 30, 2017.
|
| 258 |
+
|
| 259 |
+
Sebastian Thrun. Lifelong learning algorithms. In Learning to learn, pp. 181–209. Springer, 1998.
|
| 260 |
+
|
| 261 |
+
V.N. Vapnik. Statistical Learning Theory. Wiley, 1998. ISBN 9788126528929.
|
| 262 |
+
|
| 263 |
+
Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems 29, pp. 3630–3638, 2016.
|
| 264 |
+
|
| 265 |
+
David Yarowsky. Unsupervised word sense disambiguation rivaling supervised methods. In Proceedings of the 33rd annual meeting on Association for Computational Linguistics, pp. 189– 196. Association for Computational Linguistics, 1995.
|
| 266 |
+
|
| 267 |
+
Xiaojin Zhu. Semi-supervised learning literature survey. 2005.
|
| 268 |
+
|
| 269 |
+
# A OMNIGLOT DATASET DETAILS
|
| 270 |
+
|
| 271 |
+
We used the following split details for experiments on Omniglot dataset. This is the same train/test split as (Vinyals et al., 2016), but we created our own validation split for selecting hyper-parameters. Models are trained on the train split only.
|
| 272 |
+
|
| 273 |
+
Train Alphabets: Alphabet_of_the_Magi, Angelic, Anglo-Saxon_Futhorc, Arcadian, Asomtavruli_(Georgian), Atemayar_Qelisayer, Atlantean, Aurek-Besh, Avesta, Balinese, Blackfoot_(Canadian_Aboriginal_Syllabics), Braille, Burmese_(Myanmar), Cyrillic, Futurama, Ge_ez, Glagolitic, Grantha, Greek, Gujarati, Gurmukhi (character 01-41), Inuktitut_(Canadian_Aboriginal_Syllabics), Japanese_(hiragana), Japanese_(katakana), Korean, Latin, Malay_(Jawi_-_Arabic), N_Ko, Ojibwe_(Canadian_Aboriginal_Syllabics), Sanskrit, Syriac_(Estrangelo), Tagalog, Tifinagh
|
| 274 |
+
|
| 275 |
+
Validation Alphabets: Armenian, Bengali, Early_Aramaic, Hebrew, Mkhedruli_(Geogian)
|
| 276 |
+
|
| 277 |
+
Test Alphabets: Gurmukhi (character 42-45), Kannada, Keble, Malayalam, Manipuri, Mongolian, Old_Church_Slavonic_(Cyrillic), Oriya, Sylheti, Syriac_(Serto), Tengwar, Tibetan, ULOG
|
| 278 |
+
|
| 279 |
+
# B tieredIMAGENET DATASET DETAILS
|
| 280 |
+
|
| 281 |
+
Each high-level category in tieredImageNet contains between 10 and 30 ILSVRC-12 classes (17.8 on average). In the ImageNet hierarchy, some classes have multiple parent nodes. Therefore, classes belonging to more than one category were removed from the dataset to ensure separation between training and test categories. Test categories were chosen to reflect various levels of separation between training and test classes. Some test categories (such as “working dog”) are fairly similar to training categories, whereas others (such as “geological formation”) are quite different. The list of categories is shown below and statistics of the dataset can be found in Table 4. A visualization of the categories according to the ImageNet hierarchy is shown in Figure 5. The full list of classes per category will also be made public, however for the sake of brevity we do not include it here.
|
| 282 |
+
|
| 283 |
+
Table 4: Statistics of the tieredImageNet dataset.
|
| 284 |
+
|
| 285 |
+
<table><tr><td></td><td>Train</td><td>Val</td><td>Test</td><td>Total</td></tr><tr><td>Categories</td><td>20</td><td>6</td><td>8</td><td>34</td></tr><tr><td>Classes</td><td>351</td><td>97</td><td>160</td><td>608</td></tr><tr><td>Images</td><td>448,695</td><td>124,261</td><td>206,209</td><td>779,165</td></tr></table>
|
| 286 |
+
|
| 287 |
+
Train Categories: n02087551 (hound, hound dog), n02092468 (terrier), n02120997 (feline, felid), n02370806 (ungulate, hoofed mammal), n02469914 (primate), n01726692 (snake, serpent, ophidian), n01674216 (saurian), n01524359 (passerine, passeriform bird), n01844917 (aquatic bird), n04081844 (restraint, constraint), n03574816 (instrument), n03800933 (musical instrument, instrument), n03125870 (craft), n04451818 (tool), n03414162 (game equipment), n03278248 (electronic equipment), n03419014 (garment), n03297735 (establishment), n02913152 (building, edifice), n04014297 (protective covering, protective cover, protection).
|
| 288 |
+
|
| 289 |
+
Validation Categories: n02098550 (sporting dog, gun dog), n03257877 (durables, durable goods, consumer durables), n03405265 (furnishing), n03699975 (machine), n03738472 (mechanism), n03791235 (motor vehicle, automotive vehicle),
|
| 290 |
+
|
| 291 |
+
Test Categories: n02103406 (working dog), n01473806 (aquatic vertebrate), n02159955 (insect), n04531098 (vessel), n03839993 (obstruction, obstructor, obstructer, impediment, impedimenta), n09287968 (geological formation, formation), n00020090 (substance), n15046900 (solid).
|
| 292 |
+
|
| 293 |
+
Table 5: Few-shot learning baseline results using labeled/unlabeled splits. Baselines either takes inputs directly from the pixel space or use a CNN to extract features. “rnd” denotes using a randomly initialized CNN, and “pre” denotes using a CNN that is pretrained for supervised classification for all training classes.
|
| 294 |
+
|
| 295 |
+
<table><tr><td rowspan="2">Models</td><td>Omniglot</td><td colspan="2">miniImageNet</td><td colspan="2">tieredImageNet</td></tr><tr><td>1-shot</td><td>1-shot</td><td>5-shot</td><td>1-shot</td><td>5-shot</td></tr><tr><td>1-NN Pixel</td><td>40.39± 0.36</td><td>26.74 ± 0.48</td><td>31.43 ± 0.51</td><td>26.55± 0.50</td><td>30.79± 0.53</td></tr><tr><td>1-NN CNN rnd</td><td>59.55 ± 0.46</td><td>24.03 ± 0.38</td><td>27.54 ± 0.42</td><td>25.49 ± 0.45</td><td>30.01 ± 0.47</td></tr><tr><td>1-NN CNN pre</td><td>52.53 ± 0.51</td><td>32.90 ± 0.58</td><td>40.79 ± 0.76</td><td>32.76 ± 0.66</td><td>40.26 ± 0.67</td></tr><tr><td>LRPixel</td><td>49.15 ± 0.39</td><td>24.50 ± 0.41</td><td>33.33±0.68</td><td>25.70± 0.46</td><td>36.30± 0.62</td></tr><tr><td>LR CNN rnd</td><td>57.80 ± 0.45</td><td>24.10 ± 0.50</td><td>28.40 ± 0.42</td><td>26.55 ± 0.48</td><td>32.51 ± 0.52</td></tr><tr><td>LR CNN pre</td><td>48.49 ± 0.47</td><td>30.28 ± 0.54</td><td>40.27 ± 0.59</td><td>34.52 ± 0.68</td><td>43.58 ± 0.72</td></tr><tr><td>ProtoNet</td><td>94.62 ± 0.09</td><td>43.61 ± 0.27</td><td>59.08 ± 0.22</td><td>46.52 ± 0.32</td><td>66.15 ± 0.34</td></tr></table>
|
| 296 |
+
|
| 297 |
+
# C EXTRA EXPERIMENTAL RESULTS
|
| 298 |
+
|
| 299 |
+
# C.1 FEW-SHOT CLASSIFICATION BASELINES
|
| 300 |
+
|
| 301 |
+
We provide baseline results on few-shot classification using 1-nearest neighbor and logistic regression with either pixel inputs or CNN features. Compared with the baselines, Regular ProtoNet performs significantly better on all three few-shot classification datasets.
|
| 302 |
+
|
| 303 |
+
# C.2 NUMBER OF UNLABELED ITEMS
|
| 304 |
+
|
| 305 |
+
Figure 6 shows test accuracy values with different number of unlabeled items during test time.
|
| 306 |
+
Figure 7 shows our mask output value distribution of the Masked Soft $k$ -Means model on Omniglot.
|
| 307 |
+
The mask values have a bi-modal distribution, corresponding to distractor and non-distractor items.
|
| 308 |
+
|
| 309 |
+
# D HYPERPARAMETER DETAILS
|
| 310 |
+
|
| 311 |
+
For Omniglot, we adopted the best hyperparameter settings found for ordinary Prototypical Networks in Snell et al. (2017). In these settings, the learning rate was set to 1e-3, and cut in half every 2K updates starting at update 2K. We trained for a total of 20K updates. For miniImagenet and tieredImageNet, we trained with a starting learning rate of 1e-3, which we also decayed. We started the decay after 25K updates, and every 25K updates thereafter we cut it in half. We trained for a total of 200K updates. We used ADAM (Kingma & Ba, 2014) for the optimization of our models. For the MLP used in the Masked Soft $k$ -Means model, we use a single hidden layer with 20 hidden units with a tanh non-linearity for all 3 datasets. We did not tune the hyparameters of this MLP so better performance may be attained with a more rigorous hyperparameter search.
|
| 312 |
+
|
| 313 |
+

|
| 314 |
+
|
| 315 |
+

|
| 316 |
+
Figure 6: Model Performance on tieredImageNet with different number of unlabeled items during test time. We include test accuracy numbers in this chart.
|
| 317 |
+
|
| 318 |
+

|
| 319 |
+
Figure 7: Mask values predicted by masked soft k-means on Omniglot.
|
parse/train/HJcSzz-CZ/HJcSzz-CZ_content_list.json
ADDED
|
@@ -0,0 +1,1701 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "META-LEARNING FOR SEMI-SUPERVISED FEW-SHOTCLASSIFICATION",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
176,
|
| 8 |
+
99,
|
| 9 |
+
823,
|
| 10 |
+
145
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Mengye $\\mathbf { R e n } ^ { \\dagger \\ltimes }$ , Eleni Triantafillou∗ †on, Sachin Ravi∗ §, Jake Snell†on, Kevin Swersky¶, Joshua B. Tenenbaum\\, Hugo Larochelle¶‡ & Richard S. Zemel†‡on ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
186,
|
| 19 |
+
169,
|
| 20 |
+
782,
|
| 21 |
+
199
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "†University of Toronto, §Princeton University, ¶Google Brain, \\MIT, ‡CIFAR, onVector Institute {mren,eleni}@cs.toronto.edu, sachinr@cs.princeton.edu, jsnell@cs.toronto.edu, kswersky@google.com, jbt@mit.edu, hugolarochelle@google.com, zemel@cs.toronto.edu ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
186,
|
| 30 |
+
200,
|
| 31 |
+
810,
|
| 32 |
+
256
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "ABSTRACT ",
|
| 39 |
+
"text_level": 1,
|
| 40 |
+
"bbox": [
|
| 41 |
+
454,
|
| 42 |
+
292,
|
| 43 |
+
544,
|
| 44 |
+
308
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "In few-shot classification, we are interested in learning algorithms that train a classifier from only a handful of labeled examples. Recent progress in few-shot classification has featured meta-learning, in which a parameterized model for a learning algorithm is defined and trained on episodes representing different classification problems, each with a small labeled training set and its corresponding test set. In this work, we advance this few-shot classification paradigm towards a scenario where unlabeled examples are also available within each episode. We consider two situations: one where all unlabeled examples are assumed to belong to the same set of classes as the labeled examples of the episode, as well as the more challenging situation where examples from other distractor classes are also provided. To address this paradigm, we propose novel extensions of Prototypical Networks (Snell et al., 2017) that are augmented with the ability to use unlabeled examples when producing prototypes. These models are trained in an end-to-end way on episodes, to learn to leverage the unlabeled examples successfully. We evaluate these methods on versions of the Omniglot and miniImageNet benchmarks, adapted to this new framework augmented with unlabeled examples. We also propose a new split of ImageNet, consisting of a large set of classes, with a hierarchical structure. Our experiments confirm that our Prototypical Networks can learn to improve their predictions due to unlabeled examples, much like a semi-supervised algorithm would. ",
|
| 51 |
+
"bbox": [
|
| 52 |
+
233,
|
| 53 |
+
324,
|
| 54 |
+
764,
|
| 55 |
+
602
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "1 INTRODUCTION ",
|
| 62 |
+
"text_level": 1,
|
| 63 |
+
"bbox": [
|
| 64 |
+
176,
|
| 65 |
+
628,
|
| 66 |
+
336,
|
| 67 |
+
643
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "The availability of large quantities of labeled data has enabled deep learning methods to achieve impressive breakthroughs in several tasks related to artificial intelligence, such as speech recognition, object recognition and machine translation. However, current deep learning approaches struggle in tackling problems for which labeled data are scarce. Specifically, while current methods excel at tackling a single problem with lots of labeled data, methods that can simultaneously solve a large variety of problems that each have only a few labels are lacking. Humans on the other hand are readily able to rapidly learn new classes, such as new types of fruit when we visit a tropical country. This significant gap between human and machine learning provides fertile ground for deep learning developments. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
660,
|
| 77 |
+
825,
|
| 78 |
+
785
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "For this reason, recently there has been an increasing body of work on few-shot learning, which considers the design of learning algorithms that specifically allow for better generalization on problems with small labeled training sets. Here we focus on the case of few-shot classification, where the given classification problem is assumed to contain only a handful of labeled examples per class. One approach to few-shot learning follows a form of meta-learning 1 (Thrun, 1998; Hochreiter et al., 2001), which performs transfer learning from a pool of various classification problems generated from large quantities of available labeled data, to new classification problems from classes unseen at training time. Meta-learning may take the form of learning a shared metric (Vinyals et al., 2016; Snell et al., 2017), a common initialization for few-shot classifiers (Ravi & Larochelle, 2017; Finn et al., 2017) or a generic inference network (Santoro et al., 2016; Mishra et al., 2017). ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
173,
|
| 87 |
+
791,
|
| 88 |
+
825,
|
| 89 |
+
876
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "text",
|
| 95 |
+
"text": "",
|
| 96 |
+
"bbox": [
|
| 97 |
+
174,
|
| 98 |
+
103,
|
| 99 |
+
825,
|
| 100 |
+
159
|
| 101 |
+
],
|
| 102 |
+
"page_idx": 1
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"type": "text",
|
| 106 |
+
"text": "These various meta-learning formulations have led to significant progress recently in few-shot classification. However, this progress has been limited in the setup of each few-shot learning episode, which differs from how humans learn new concepts in many dimensions. In this paper we aim to generalize the setup in two ways. First, we consider a scenario where the new classes are learned in the presence of additional unlabeled data. While there have been many successful applications of semisupervised learning to the regular setting of a single classification task (Chapelle et al., 2010) where classes at training and test time are the same, such work has not addressed the challenge of performing transfer to new classes never seen at training time, which we consider here. Second, we consider the situation where the new classes to be learned are not viewed in isolation. Instead, many of the unlabeled examples are from different classes; the presence of such distractor classes introduces an additional and more realistic level of difficulty to the fewshot problem. ",
|
| 107 |
+
"bbox": [
|
| 108 |
+
174,
|
| 109 |
+
166,
|
| 110 |
+
483,
|
| 111 |
+
498
|
| 112 |
+
],
|
| 113 |
+
"page_idx": 1
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"type": "image",
|
| 117 |
+
"img_path": "images/712ced75c31849892a977ed5de7faabcb52f69269afed899f0f80eac77b18e73.jpg",
|
| 118 |
+
"image_caption": [
|
| 119 |
+
"Figure 1: Consider a setup where the aim is to learn a classifier to distinguish between two previously unseen classes, goldfish and shark, given not only labeled examples of these two classes, but also a larger pool of unlabeled examples, some of which may belong to one of these two classes of interest. In this work we aim to move a step closer to this more natural learning framework by incorporating in our learning episodes unlabeled data from the classes we aim to learn representations for (shown with dashed red borders) as well as from distractor classes . "
|
| 120 |
+
],
|
| 121 |
+
"image_footnote": [],
|
| 122 |
+
"bbox": [
|
| 123 |
+
504,
|
| 124 |
+
184,
|
| 125 |
+
815,
|
| 126 |
+
332
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "This work is a first study of this challenging semi-supervised form of few-shot learning. First, we define the problem and propose benchmarks for evaluation that are adapted from the Omniglot and miniImageNet benchmarks used in ordinary few-shot learning. We perform an extensive empirical investigation of the two settings mentioned above, with and without distractor classes. Second, we propose and study three novel extensions of Prototypical Networks (Snell et al., 2017), a state-ofthe-art approach to few-shot learning, to the semi-supervised setting. Finally, we demonstrate in our experiments that our semi-supervised variants successfully learn to leverage unlabeled examples and outperform purely supervised Prototypical Networks. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
506,
|
| 136 |
+
825,
|
| 137 |
+
617
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "2 BACKGROUND ",
|
| 144 |
+
"text_level": 1,
|
| 145 |
+
"bbox": [
|
| 146 |
+
176,
|
| 147 |
+
640,
|
| 148 |
+
326,
|
| 149 |
+
656
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 1
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "text",
|
| 155 |
+
"text": "We start by defining precisely the current paradigm for few-shot learning and the Prototypical Network approach to this problem. ",
|
| 156 |
+
"bbox": [
|
| 157 |
+
176,
|
| 158 |
+
672,
|
| 159 |
+
825,
|
| 160 |
+
702
|
| 161 |
+
],
|
| 162 |
+
"page_idx": 1
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"type": "text",
|
| 166 |
+
"text": "2.1 FEW-SHOT LEARNING ",
|
| 167 |
+
"text_level": 1,
|
| 168 |
+
"bbox": [
|
| 169 |
+
176,
|
| 170 |
+
722,
|
| 171 |
+
364,
|
| 172 |
+
734
|
| 173 |
+
],
|
| 174 |
+
"page_idx": 1
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"type": "text",
|
| 178 |
+
"text": "Recent progress on few-shot learning has been made possible by following an episodic paradigm. Consider a situation where we have a large labeled dataset for a set of classes $\\mathcal { C } _ { \\mathrm { t r a i n } }$ . However, after training on examples from $\\mathcal { C } _ { \\mathrm { t r a i n } }$ , our ultimate goal is to produce classifiers for a disjoint set of new classes $\\mathcal { C } _ { \\mathrm { t e s t } }$ , for which only a few labeled examples will be available. The idea behind the episodic paradigm is to simulate the types of few-shot problems that will be encountered at test, taking advantage of the large quantities of available labeled data for classes $\\mathcal { C } _ { \\mathrm { t r a i n } }$ . ",
|
| 179 |
+
"bbox": [
|
| 180 |
+
174,
|
| 181 |
+
747,
|
| 182 |
+
825,
|
| 183 |
+
832
|
| 184 |
+
],
|
| 185 |
+
"page_idx": 1
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"type": "text",
|
| 189 |
+
"text": "Specifically, models are trained on $K$ -shot, $N$ -way episodes constructed by first sampling a small subset of $N$ classes from $\\mathcal { C } _ { \\mathrm { t r a i n } }$ and then generating: 1) a training (support) set ${ \\boldsymbol { s } } \\ =$ $\\left\\{ ( \\pmb { x } _ { 1 } , y _ { 1 } ) , ( \\pmb { x } _ { 2 } , y _ { 2 } ) , \\dots , ( \\pmb { x } _ { N \\times K } , y _ { N \\times K } ) \\right\\}$ containing $K$ examples from each of the $N$ classes and 2) a test (query) set $\\mathcal { Q } = \\{ ( \\pmb { x } _ { 1 } ^ { * } , y _ { 1 } ^ { * } ) , ( \\pmb { x } _ { 2 } ^ { * } , y _ { 2 } ^ { * } ) , . . . , ( \\pmb { x } _ { T } ^ { * } , y _ { T } ^ { * } ) \\}$ of different examples from the same $N$ classes. Each $\\pmb { x } _ { i } \\in \\mathbb { R } ^ { D }$ is an input vector of dimension $D$ and $y _ { i } \\in \\{ 1 , 2 , \\ldots , N \\}$ is a class label (similarly for $\\boldsymbol { \\mathscr { x } } _ { i } ^ { * }$ and $y _ { i } ^ { * }$ ). Training on such episodes is done by feeding the support set $s$ to the model and updating its parameters to minimize the loss of its predictions for the examples in the query set $\\mathcal { Q }$ . ",
|
| 190 |
+
"bbox": [
|
| 191 |
+
174,
|
| 192 |
+
838,
|
| 193 |
+
825,
|
| 194 |
+
924
|
| 195 |
+
],
|
| 196 |
+
"page_idx": 1
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"type": "text",
|
| 200 |
+
"text": "",
|
| 201 |
+
"bbox": [
|
| 202 |
+
173,
|
| 203 |
+
103,
|
| 204 |
+
823,
|
| 205 |
+
132
|
| 206 |
+
],
|
| 207 |
+
"page_idx": 2
|
| 208 |
+
},
|
| 209 |
+
{
|
| 210 |
+
"type": "text",
|
| 211 |
+
"text": "One way to think of this approach is that our model effectively trains to be a good learning algorithm. Indeed, much like a learning algorithm, the model must take in a set of labeled examples and produce a predictor that can be applied to new examples. Moreover, training directly encourages the classifier produced by the model to have good generalization on the new examples of the query set. Due to this analogy, training under this paradigm is often referred to as learning to learn or meta-learning. ",
|
| 212 |
+
"bbox": [
|
| 213 |
+
174,
|
| 214 |
+
138,
|
| 215 |
+
825,
|
| 216 |
+
209
|
| 217 |
+
],
|
| 218 |
+
"page_idx": 2
|
| 219 |
+
},
|
| 220 |
+
{
|
| 221 |
+
"type": "text",
|
| 222 |
+
"text": "On the other hand, referring to the content of episodes as training and test sets and to the process of learning on these episodes as meta-learning or meta-training (as is sometimes done in the literature) can be confusing. So for the sake of clarity, we will refer to the content of episodes as support and query sets, and to the process of iterating over the training episodes simply as training. ",
|
| 223 |
+
"bbox": [
|
| 224 |
+
174,
|
| 225 |
+
215,
|
| 226 |
+
825,
|
| 227 |
+
272
|
| 228 |
+
],
|
| 229 |
+
"page_idx": 2
|
| 230 |
+
},
|
| 231 |
+
{
|
| 232 |
+
"type": "text",
|
| 233 |
+
"text": "2.2 PROTOTYPICAL NETWORKS ",
|
| 234 |
+
"text_level": 1,
|
| 235 |
+
"bbox": [
|
| 236 |
+
176,
|
| 237 |
+
290,
|
| 238 |
+
406,
|
| 239 |
+
305
|
| 240 |
+
],
|
| 241 |
+
"page_idx": 2
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"type": "text",
|
| 245 |
+
"text": "Prototypical Network (Snell et al., 2017) is a few-shot learning model that has the virtue of being simple and yet obtaining state-of-the-art performance. At a high-level, it uses the support set $s$ to extract a prototype vector from each class, and classifies the inputs in the query set based on their distance to the prototype of each class. ",
|
| 246 |
+
"bbox": [
|
| 247 |
+
174,
|
| 248 |
+
316,
|
| 249 |
+
825,
|
| 250 |
+
373
|
| 251 |
+
],
|
| 252 |
+
"page_idx": 2
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"type": "text",
|
| 256 |
+
"text": "More precisely, Prototypical Networks learn an embedding function $h ( { \\pmb x } )$ , parameterized as a neural network, that maps examples into a space where examples from the same class are close and those from different classes are far. All parameters of Prototypical Networks lie in the embedding function. ",
|
| 257 |
+
"bbox": [
|
| 258 |
+
174,
|
| 259 |
+
381,
|
| 260 |
+
825,
|
| 261 |
+
422
|
| 262 |
+
],
|
| 263 |
+
"page_idx": 2
|
| 264 |
+
},
|
| 265 |
+
{
|
| 266 |
+
"type": "text",
|
| 267 |
+
"text": "To compute the prototype $\\pmb { p } _ { c }$ of each class $c$ , a per-class average of the embedded examples is performed: ",
|
| 268 |
+
"bbox": [
|
| 269 |
+
171,
|
| 270 |
+
429,
|
| 271 |
+
821,
|
| 272 |
+
458
|
| 273 |
+
],
|
| 274 |
+
"page_idx": 2
|
| 275 |
+
},
|
| 276 |
+
{
|
| 277 |
+
"type": "equation",
|
| 278 |
+
"img_path": "images/9e85294a78e37f1b293004b973076ecc152ad8fd4fed17242c7977b8470aa8e0.jpg",
|
| 279 |
+
"text": "$$\np _ { c } = \\frac { \\sum _ { i } h ( \\pmb { x } _ { i } ) z _ { i , c } } { \\sum _ { i } z _ { i , c } } , \\mathrm { w h e r e } z _ { i , c } = \\mathbb { 1 } [ y _ { i } = c ] .\n$$",
|
| 280 |
+
"text_format": "latex",
|
| 281 |
+
"bbox": [
|
| 282 |
+
343,
|
| 283 |
+
467,
|
| 284 |
+
650,
|
| 285 |
+
502
|
| 286 |
+
],
|
| 287 |
+
"page_idx": 2
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"type": "text",
|
| 291 |
+
"text": "These prototypes define a predictor for the class of any new (query) example $\\mathbf { \\boldsymbol { x } } ^ { * }$ , which assigns a probability over any class $c$ based on the distances between $\\pmb { x } ^ { * }$ and each prototype, as follows: ",
|
| 292 |
+
"bbox": [
|
| 293 |
+
174,
|
| 294 |
+
510,
|
| 295 |
+
826,
|
| 296 |
+
539
|
| 297 |
+
],
|
| 298 |
+
"page_idx": 2
|
| 299 |
+
},
|
| 300 |
+
{
|
| 301 |
+
"type": "equation",
|
| 302 |
+
"img_path": "images/c67f8b15c0114356f27f27eb255d5b4a343fda7f2c04d142107772305407271f.jpg",
|
| 303 |
+
"text": "$$\np ( c | \\pmb { x } ^ { * } , \\{ p _ { c } \\} ) = \\frac { \\exp ( - | | h ( \\pmb { x } ^ { * } ) - \\pmb { p } _ { c } | | _ { 2 } ^ { 2 } ) } { \\sum _ { c ^ { \\prime } } \\exp ( - | | h ( \\pmb { x } ^ { * } ) - \\pmb { p } _ { c ^ { \\prime } } | | _ { 2 } ^ { 2 } ) } .\n$$",
|
| 304 |
+
"text_format": "latex",
|
| 305 |
+
"bbox": [
|
| 306 |
+
339,
|
| 307 |
+
546,
|
| 308 |
+
656,
|
| 309 |
+
583
|
| 310 |
+
],
|
| 311 |
+
"page_idx": 2
|
| 312 |
+
},
|
| 313 |
+
{
|
| 314 |
+
"type": "text",
|
| 315 |
+
"text": "The loss function used to update Prototypical Networks for a given training episode is then simply the average negative log-probability of the correct class assignments, for all query examples: ",
|
| 316 |
+
"bbox": [
|
| 317 |
+
173,
|
| 318 |
+
589,
|
| 319 |
+
821,
|
| 320 |
+
619
|
| 321 |
+
],
|
| 322 |
+
"page_idx": 2
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"type": "equation",
|
| 326 |
+
"img_path": "images/517e9483d097dc2911216a91e829f6af22eecf49ecdcafd4ce1072ca3d470d6f.jpg",
|
| 327 |
+
"text": "$$\n- \\frac { 1 } { T } \\sum _ { i } \\log p ( y _ { i } ^ { * } | \\pmb { x } _ { i } ^ { * } , \\{ \\pmb { p } _ { c } \\} ) .\n$$",
|
| 328 |
+
"text_format": "latex",
|
| 329 |
+
"bbox": [
|
| 330 |
+
398,
|
| 331 |
+
627,
|
| 332 |
+
596,
|
| 333 |
+
664
|
| 334 |
+
],
|
| 335 |
+
"page_idx": 2
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"type": "text",
|
| 339 |
+
"text": "Training proceeds by minimizing the average loss, iterating over training episodes and performing a gradient descent update for each. ",
|
| 340 |
+
"bbox": [
|
| 341 |
+
174,
|
| 342 |
+
671,
|
| 343 |
+
823,
|
| 344 |
+
700
|
| 345 |
+
],
|
| 346 |
+
"page_idx": 2
|
| 347 |
+
},
|
| 348 |
+
{
|
| 349 |
+
"type": "text",
|
| 350 |
+
"text": "Generalization performance is measured on test set episodes, which contain images from classes in $\\mathcal { C } _ { \\mathrm { t e s t } }$ instead of $\\mathcal { C } _ { \\mathrm { t r a i n } }$ . For each test episode, we use the predictor produced by the Prototypical Network for the provided support set $s$ to classify each of query input $\\pmb { x } ^ { * }$ into the most likely class $\\hat { y } = \\operatorname { a r g m a x } _ { c } p ( \\bar { c } | \\pmb { x } ^ { * } , \\{ \\pmb { p } _ { c } \\} )$ . ",
|
| 351 |
+
"bbox": [
|
| 352 |
+
174,
|
| 353 |
+
707,
|
| 354 |
+
825,
|
| 355 |
+
765
|
| 356 |
+
],
|
| 357 |
+
"page_idx": 2
|
| 358 |
+
},
|
| 359 |
+
{
|
| 360 |
+
"type": "text",
|
| 361 |
+
"text": "3 SEMI-SUPERVISED FEW-SHOT LEARNING ",
|
| 362 |
+
"text_level": 1,
|
| 363 |
+
"bbox": [
|
| 364 |
+
174,
|
| 365 |
+
785,
|
| 366 |
+
553,
|
| 367 |
+
801
|
| 368 |
+
],
|
| 369 |
+
"page_idx": 2
|
| 370 |
+
},
|
| 371 |
+
{
|
| 372 |
+
"type": "text",
|
| 373 |
+
"text": "We now define the semi-supervised setting considered in this work for few-shot learning. ",
|
| 374 |
+
"bbox": [
|
| 375 |
+
171,
|
| 376 |
+
818,
|
| 377 |
+
754,
|
| 378 |
+
833
|
| 379 |
+
],
|
| 380 |
+
"page_idx": 2
|
| 381 |
+
},
|
| 382 |
+
{
|
| 383 |
+
"type": "text",
|
| 384 |
+
"text": "The training set is denoted as a tuple of labeled and unlabeled examples: $( S , R )$ . The labeled portion is the usual support set $s$ of the few-shot learning literature, containing a list of tuples of inputs and targets. In addition to classic few-shot learning, we introduce an unlabeled set $\\mathcal { R }$ containing only inputs: $\\mathcal { R } = \\{ \\tilde { { \\pmb { x } } } _ { 1 } , \\tilde { { \\pmb { x } } } _ { 2 } , \\ldots , \\tilde { { \\pmb { x } } } _ { M } \\}$ . As in the purely supervised setting, our models are trained to perform well when predicting the labels for the examples in the episode’s query set $\\mathcal { Q }$ . Figure 2 shows a visualization of training and test episodes. ",
|
| 385 |
+
"bbox": [
|
| 386 |
+
174,
|
| 387 |
+
839,
|
| 388 |
+
825,
|
| 389 |
+
924
|
| 390 |
+
],
|
| 391 |
+
"page_idx": 2
|
| 392 |
+
},
|
| 393 |
+
{
|
| 394 |
+
"type": "image",
|
| 395 |
+
"img_path": "images/5436151688cbedebe02d2802538405afd5470e7719eccd33d5f925b943abe901.jpg",
|
| 396 |
+
"image_caption": [
|
| 397 |
+
"Figure 2: Example of the semi-supervised few-shot learning setup. Training involves iterating through training episodes, consisting of a support set $s$ , an unlabeled set $\\mathcal { R }$ , and a query set $\\mathcal { Q }$ . The goal is to use the labeled items (shown with their numeric class label) in $s$ and the unlabeled items in $\\mathcal { R }$ within each episode to generalize to good performance on the corresponding query set. The unlabeled items in $\\mathcal { R }$ may either be pertinent to the classes we are considering (shown above with green plus signs) or they may be distractor items which belong to a class that is not relevant to the current episode (shown with red minus signs). However note that the model does not actually have ground truth information as to whether each unlabeled example is a distractor or not; the plus/minus signs are shown only for illustrative purposes. At test time, we are given new episodes consisting of novel classes not seen during training that we use to evaluate the meta-learning method. "
|
| 398 |
+
],
|
| 399 |
+
"image_footnote": [],
|
| 400 |
+
"bbox": [
|
| 401 |
+
250,
|
| 402 |
+
102,
|
| 403 |
+
736,
|
| 404 |
+
342
|
| 405 |
+
],
|
| 406 |
+
"page_idx": 3
|
| 407 |
+
},
|
| 408 |
+
{
|
| 409 |
+
"type": "text",
|
| 410 |
+
"text": "3.1 SEMI-SUPERVISED PROTOTYPICAL NETWORKS ",
|
| 411 |
+
"text_level": 1,
|
| 412 |
+
"bbox": [
|
| 413 |
+
174,
|
| 414 |
+
498,
|
| 415 |
+
542,
|
| 416 |
+
512
|
| 417 |
+
],
|
| 418 |
+
"page_idx": 3
|
| 419 |
+
},
|
| 420 |
+
{
|
| 421 |
+
"type": "text",
|
| 422 |
+
"text": "In their original formulation, Prototypical Networks do not specify a way to leverage the unlabeled set $\\mathcal { R }$ . In what follows, we now propose various extensions that start from the basic definition of prototypes $\\pmb { p _ { c } }$ and provide a procedure for producing refined prototypes $\\tilde { p } _ { c }$ using the unlabeled examples in $\\mathcal { R }$ . ",
|
| 423 |
+
"bbox": [
|
| 424 |
+
174,
|
| 425 |
+
525,
|
| 426 |
+
825,
|
| 427 |
+
580
|
| 428 |
+
],
|
| 429 |
+
"page_idx": 3
|
| 430 |
+
},
|
| 431 |
+
{
|
| 432 |
+
"type": "text",
|
| 433 |
+
"text": "After the refined prototypes are obtained, each of these models is trained with the same loss function for ordinary Prototypical Networks of Equation 3, but replacing ${ \\pmb p } _ { c }$ with $\\tilde { p } _ { c }$ . That is, each query example is classified into one of the $N$ classes based on the proximity of its embedded position with the corresponding refined prototypes, and the average negative logprobability of the correct classification is used for training. ",
|
| 434 |
+
"bbox": [
|
| 435 |
+
174,
|
| 436 |
+
588,
|
| 437 |
+
387,
|
| 438 |
+
781
|
| 439 |
+
],
|
| 440 |
+
"page_idx": 3
|
| 441 |
+
},
|
| 442 |
+
{
|
| 443 |
+
"type": "image",
|
| 444 |
+
"img_path": "images/5e1bc913e73854521429d9147f4e5fea101390e46ff01bf72a7b6a79261c8160.jpg",
|
| 445 |
+
"image_caption": [
|
| 446 |
+
"Figure 3: Left: The prototypes are initialized based on the mean location of the examples of the corresponding class, as in ordinary Prototypical Networks. Support, unlabeled, and query examples have solid, dashed, and white colored borders respectively. Right: The refined prototypes obtained by incorporating the unlabeled examples, which classifies all query examples correctly. "
|
| 447 |
+
],
|
| 448 |
+
"image_footnote": [],
|
| 449 |
+
"bbox": [
|
| 450 |
+
416,
|
| 451 |
+
594,
|
| 452 |
+
808,
|
| 453 |
+
689
|
| 454 |
+
],
|
| 455 |
+
"page_idx": 3
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"type": "text",
|
| 459 |
+
"text": "3.1.1 PROTOTYPICAL NETWORKS WITH SOFT $k$ -MEANS ",
|
| 460 |
+
"text_level": 1,
|
| 461 |
+
"bbox": [
|
| 462 |
+
174,
|
| 463 |
+
801,
|
| 464 |
+
575,
|
| 465 |
+
815
|
| 466 |
+
],
|
| 467 |
+
"page_idx": 3
|
| 468 |
+
},
|
| 469 |
+
{
|
| 470 |
+
"type": "text",
|
| 471 |
+
"text": "We first consider a simple way of leveraging unlabeled examples for refining prototypes, by taking inspiration from semi-supervised clustering. Viewing each prototype as a cluster center, the refinement process could attempt to adjust the cluster locations to better fit the examples in both the support and unlabeled sets. Under this view, cluster assignments of the labeled examples in the support set are considered known and fixed to each example’s label. The refinement process must instead estimate the cluster assignments of the unlabeled examples and adjust the cluster locations (the prototypes) accordingly. ",
|
| 472 |
+
"bbox": [
|
| 473 |
+
174,
|
| 474 |
+
827,
|
| 475 |
+
825,
|
| 476 |
+
924
|
| 477 |
+
],
|
| 478 |
+
"page_idx": 3
|
| 479 |
+
},
|
| 480 |
+
{
|
| 481 |
+
"type": "text",
|
| 482 |
+
"text": "One natural choice would be to borrow from the inference performed by soft $k$ -means. We prefer this version of $k$ -means over hard assignments since hard assignments would make the inference non-differentiable. We start with the regular Prototypical Network’s prototypes $\\pmb { p _ { c } }$ (as specified in Equation 1) as the cluster locations. Then, the unlabeled examples get a partial assignment $( \\tilde { z } _ { j , c } )$ to each cluster based on their Euclidean distance to the cluster locations. Finally, refined prototypes are obtained by incorporating these unlabeled examples. ",
|
| 483 |
+
"bbox": [
|
| 484 |
+
174,
|
| 485 |
+
103,
|
| 486 |
+
825,
|
| 487 |
+
188
|
| 488 |
+
],
|
| 489 |
+
"page_idx": 4
|
| 490 |
+
},
|
| 491 |
+
{
|
| 492 |
+
"type": "text",
|
| 493 |
+
"text": "This process can be summarized as follows: ",
|
| 494 |
+
"bbox": [
|
| 495 |
+
174,
|
| 496 |
+
194,
|
| 497 |
+
464,
|
| 498 |
+
208
|
| 499 |
+
],
|
| 500 |
+
"page_idx": 4
|
| 501 |
+
},
|
| 502 |
+
{
|
| 503 |
+
"type": "equation",
|
| 504 |
+
"img_path": "images/2d9af4f0f1046794e3da41d25e05193e5a7ed20b32de7a87702a1961793b02fe.jpg",
|
| 505 |
+
"text": "$$\n\\tilde { p } _ { c } = \\frac { \\sum _ { i } h ( { \\bf x } _ { i } ) z _ { i , c } + \\sum _ { j } h ( \\tilde { x } _ { j } ) \\tilde { z } _ { j , c } } { \\sum _ { i } z _ { i , c } + \\sum _ { j } \\tilde { z } _ { j , c } } , \\mathrm { w h e r e } z _ { j , c } = \\frac { \\exp \\left( - | | h ( \\tilde { x } _ { j } ) - p _ { c } | | _ { 2 } ^ { 2 } \\right) } { \\sum _ { c ^ { \\prime } } \\exp \\left( - | | h ( \\tilde { x } _ { j } ) - p _ { c ^ { \\prime } } | | _ { 2 } ^ { 2 } \\right) }\n$$",
|
| 506 |
+
"text_format": "latex",
|
| 507 |
+
"bbox": [
|
| 508 |
+
225,
|
| 509 |
+
215,
|
| 510 |
+
772,
|
| 511 |
+
255
|
| 512 |
+
],
|
| 513 |
+
"page_idx": 4
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"type": "text",
|
| 517 |
+
"text": "Predictions of each query input’s class is then modeled as in Equation 2, but using the refined prototypes $\\tilde { p } _ { c }$ . ",
|
| 518 |
+
"bbox": [
|
| 519 |
+
174,
|
| 520 |
+
262,
|
| 521 |
+
825,
|
| 522 |
+
291
|
| 523 |
+
],
|
| 524 |
+
"page_idx": 4
|
| 525 |
+
},
|
| 526 |
+
{
|
| 527 |
+
"type": "text",
|
| 528 |
+
"text": "We could perform several iterations of refinement, as is usual in $k$ -means. However, we have experimented with various number of iterations and found results to not improve beyond a single refinement step. ",
|
| 529 |
+
"bbox": [
|
| 530 |
+
174,
|
| 531 |
+
296,
|
| 532 |
+
825,
|
| 533 |
+
340
|
| 534 |
+
],
|
| 535 |
+
"page_idx": 4
|
| 536 |
+
},
|
| 537 |
+
{
|
| 538 |
+
"type": "text",
|
| 539 |
+
"text": "3.1.2 PROTOTYPICAL NETWORKS WITH SOFT $k$ -MEANS WITH A DISTRACTOR CLUSTER ",
|
| 540 |
+
"text_level": 1,
|
| 541 |
+
"bbox": [
|
| 542 |
+
173,
|
| 543 |
+
354,
|
| 544 |
+
795,
|
| 545 |
+
371
|
| 546 |
+
],
|
| 547 |
+
"page_idx": 4
|
| 548 |
+
},
|
| 549 |
+
{
|
| 550 |
+
"type": "text",
|
| 551 |
+
"text": "The soft $k$ -means approach described above implicitly assumes that each unlabeled example belongs to either one of the $N$ classes in the episode. However, it would be much more general to not make that assumption and have a model robust to the existence of examples from other classes, which we refer to as distractor classes. For example, such a situation would arise if we wanted to distinguish between pictures of unicycles and scooters, and decided to add an unlabeled set by downloading images from the web. It then would not be realistic to assume that all these images are of unicycles or scooters. Even with a focused search, some may be from similar classes, such as bicycle. ",
|
| 552 |
+
"bbox": [
|
| 553 |
+
173,
|
| 554 |
+
380,
|
| 555 |
+
825,
|
| 556 |
+
478
|
| 557 |
+
],
|
| 558 |
+
"page_idx": 4
|
| 559 |
+
},
|
| 560 |
+
{
|
| 561 |
+
"type": "text",
|
| 562 |
+
"text": "Since soft $k$ -means distributes its soft assignments across all classes, distractor items could be harmful and interfere with the refinement process, as prototypes would be adjusted to also partially account for these distractors. A simple way to address this is to add an additional cluster whose purpose is to capture the distractors, thus preventing them from polluting the clusters of the classes of interest: ",
|
| 563 |
+
"bbox": [
|
| 564 |
+
173,
|
| 565 |
+
484,
|
| 566 |
+
825,
|
| 567 |
+
554
|
| 568 |
+
],
|
| 569 |
+
"page_idx": 4
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"type": "equation",
|
| 573 |
+
"img_path": "images/cdb2bf0f559a54d96be8ec58d46b1003e42dbd90c8e440d82a4d70751fd1569c.jpg",
|
| 574 |
+
"text": "$$\n\\pmb { p _ { c } } = \\left\\{ \\begin{array} { l l } { \\frac { \\sum _ { i } h ( \\pmb { x _ { i } } ) z _ { i , c } } { \\sum _ { i } z _ { i , c } } } & { \\quad \\mathrm { f o r } c = 1 . . . N } \\\\ { \\mathbf { 0 } } & { \\quad \\mathrm { f o r } c = N + 1 } \\end{array} \\right.\n$$",
|
| 575 |
+
"text_format": "latex",
|
| 576 |
+
"bbox": [
|
| 577 |
+
369,
|
| 578 |
+
551,
|
| 579 |
+
627,
|
| 580 |
+
594
|
| 581 |
+
],
|
| 582 |
+
"page_idx": 4
|
| 583 |
+
},
|
| 584 |
+
{
|
| 585 |
+
"type": "text",
|
| 586 |
+
"text": "Here we take the simplifying assumption that the distractor cluster has a prototype centered at the origin. We also consider introducing length-scales $r _ { c }$ to represent variations in the within-cluster distances, specifically for the distractor cluster: ",
|
| 587 |
+
"bbox": [
|
| 588 |
+
176,
|
| 589 |
+
598,
|
| 590 |
+
823,
|
| 591 |
+
640
|
| 592 |
+
],
|
| 593 |
+
"page_idx": 4
|
| 594 |
+
},
|
| 595 |
+
{
|
| 596 |
+
"type": "equation",
|
| 597 |
+
"img_path": "images/cf1189b3d3ab7be583adcc55c4841e8e627a587d9cb95cc3ef7031cedc89187d.jpg",
|
| 598 |
+
"text": "$$\n\\tilde { z } _ { j , c } = \\frac { \\exp \\left( - \\frac { 1 } { r _ { c } ^ { 2 } } | | \\tilde { x } _ { j } - p _ { c } | | _ { 2 } ^ { 2 } - A ( r _ { c } ) \\right) } { \\sum _ { c ^ { \\prime } } \\exp \\left( - \\frac { 1 } { r _ { c } ^ { 2 } } | | \\tilde { x } _ { j } - p _ { c ^ { \\prime } } | | _ { 2 } ^ { 2 } - A ( r _ { c ^ { \\prime } } ) \\right) } , \\mathrm { ~ w h e r e ~ } A ( r ) = \\frac { 1 } { 2 } \\log ( 2 \\pi ) + \\log ( r )\n$$",
|
| 599 |
+
"text_format": "latex",
|
| 600 |
+
"bbox": [
|
| 601 |
+
220,
|
| 602 |
+
646,
|
| 603 |
+
777,
|
| 604 |
+
702
|
| 605 |
+
],
|
| 606 |
+
"page_idx": 4
|
| 607 |
+
},
|
| 608 |
+
{
|
| 609 |
+
"type": "text",
|
| 610 |
+
"text": "For simplicity, we set $r _ { 1 \\dots N }$ to 1 in our experiments, and only learn the length-scale of the distractor cluster rN+1. ",
|
| 611 |
+
"bbox": [
|
| 612 |
+
174,
|
| 613 |
+
707,
|
| 614 |
+
825,
|
| 615 |
+
737
|
| 616 |
+
],
|
| 617 |
+
"page_idx": 4
|
| 618 |
+
},
|
| 619 |
+
{
|
| 620 |
+
"type": "text",
|
| 621 |
+
"text": "3.1.3 PROTOTYPICAL NETWORKS WITH SOFT $k$ -MEANS AND MASKING ",
|
| 622 |
+
"text_level": 1,
|
| 623 |
+
"bbox": [
|
| 624 |
+
174,
|
| 625 |
+
751,
|
| 626 |
+
684,
|
| 627 |
+
767
|
| 628 |
+
],
|
| 629 |
+
"page_idx": 4
|
| 630 |
+
},
|
| 631 |
+
{
|
| 632 |
+
"type": "text",
|
| 633 |
+
"text": "Modeling distractor unlabeled examples with a single cluster is likely too simplistic. Indeed, it is inconsistent with our assumption that each cluster corresponds to one class, since distractor examples may very well cover more than a single natural object category. Continuing with our unicycles and bicycles example, our web search for unlabeled images could accidentally include not only bicycles, but other related objects such as tricycles or cars. This was also reflected in our experiments, where we constructed the episode generating process so that it would sample distractor examples from multiple classes. ",
|
| 634 |
+
"bbox": [
|
| 635 |
+
173,
|
| 636 |
+
776,
|
| 637 |
+
825,
|
| 638 |
+
875
|
| 639 |
+
],
|
| 640 |
+
"page_idx": 4
|
| 641 |
+
},
|
| 642 |
+
{
|
| 643 |
+
"type": "text",
|
| 644 |
+
"text": "To address this problem, we propose an improved variant: instead of capturing distractors with a high-variance catch-all cluster, we model distractors as examples that are not within some area of any of the legitimate class prototypes. This is done by incorporating a soft-masking mechanism on the contribution of unlabeled examples. At a high level, we want unlabeled examples that are closer to a prototype to be masked less than those that are farther. ",
|
| 645 |
+
"bbox": [
|
| 646 |
+
174,
|
| 647 |
+
882,
|
| 648 |
+
825,
|
| 649 |
+
924
|
| 650 |
+
],
|
| 651 |
+
"page_idx": 4
|
| 652 |
+
},
|
| 653 |
+
{
|
| 654 |
+
"type": "text",
|
| 655 |
+
"text": "",
|
| 656 |
+
"bbox": [
|
| 657 |
+
171,
|
| 658 |
+
103,
|
| 659 |
+
823,
|
| 660 |
+
132
|
| 661 |
+
],
|
| 662 |
+
"page_idx": 5
|
| 663 |
+
},
|
| 664 |
+
{
|
| 665 |
+
"type": "text",
|
| 666 |
+
"text": "More specifically, we modify the soft $k$ -means refinement as follows. We start by computing normalized distances $\\tilde { d } _ { j , c }$ between examples $\\tilde { \\mathbfit { x } } _ { j }$ and prototypes $\\pmb { p _ { c } }$ : ",
|
| 667 |
+
"bbox": [
|
| 668 |
+
171,
|
| 669 |
+
138,
|
| 670 |
+
823,
|
| 671 |
+
170
|
| 672 |
+
],
|
| 673 |
+
"page_idx": 5
|
| 674 |
+
},
|
| 675 |
+
{
|
| 676 |
+
"type": "equation",
|
| 677 |
+
"img_path": "images/c89720438c82f2a2369fa4d7c931acee29d6be9cd4b14d7309aeb0fc03e98c41.jpg",
|
| 678 |
+
"text": "$$\n\\tilde { d } _ { j , c } = \\frac { d _ { j , c } } { \\frac { 1 } { M } \\sum _ { j } d _ { j , c } } , \\ \\mathrm { w h e r e } \\ d _ { j , c } = | | h ( \\tilde { \\bf x } _ { j } ) - { \\bf p } _ { c } | | _ { 2 } ^ { 2 }\n$$",
|
| 679 |
+
"text_format": "latex",
|
| 680 |
+
"bbox": [
|
| 681 |
+
331,
|
| 682 |
+
174,
|
| 683 |
+
666,
|
| 684 |
+
210
|
| 685 |
+
],
|
| 686 |
+
"page_idx": 5
|
| 687 |
+
},
|
| 688 |
+
{
|
| 689 |
+
"type": "text",
|
| 690 |
+
"text": "Then, soft thresholds $\\beta _ { c }$ and slopes $\\gamma _ { c }$ are predicted for each prototype, by feeding to a small neural network various statistics of the normalized distances for the prototype: ",
|
| 691 |
+
"bbox": [
|
| 692 |
+
173,
|
| 693 |
+
215,
|
| 694 |
+
821,
|
| 695 |
+
244
|
| 696 |
+
],
|
| 697 |
+
"page_idx": 5
|
| 698 |
+
},
|
| 699 |
+
{
|
| 700 |
+
"type": "equation",
|
| 701 |
+
"img_path": "images/595c570ba5f224e14933fd02aec8420606313ac0c19a50aa9d28c2e3b05349fc.jpg",
|
| 702 |
+
"text": "$$\n[ \\beta _ { c } , \\gamma _ { c } ] = \\mathrm { M L P } \\left( \\left[ \\operatorname* { m i n } _ { j } ( \\tilde { d } _ { j , c } ) , \\operatorname* { m a x } _ { j } ( \\tilde { d } _ { j , c } ) , \\operatorname { v a r } _ { j } ( \\tilde { d } _ { j , c } ) , \\operatorname { s k e w } ( \\tilde { d } _ { j , c } ) , \\operatorname { k u r t } _ { j } ( \\tilde { d } _ { j , c } ) \\right] \\right)\n$$",
|
| 703 |
+
"text_format": "latex",
|
| 704 |
+
"bbox": [
|
| 705 |
+
243,
|
| 706 |
+
247,
|
| 707 |
+
754,
|
| 708 |
+
282
|
| 709 |
+
],
|
| 710 |
+
"page_idx": 5
|
| 711 |
+
},
|
| 712 |
+
{
|
| 713 |
+
"type": "text",
|
| 714 |
+
"text": "This allows each threshold to use information on the amount of intra-cluster variation to determine how aggressively it should cut out unlabeled examples. ",
|
| 715 |
+
"bbox": [
|
| 716 |
+
173,
|
| 717 |
+
286,
|
| 718 |
+
825,
|
| 719 |
+
315
|
| 720 |
+
],
|
| 721 |
+
"page_idx": 5
|
| 722 |
+
},
|
| 723 |
+
{
|
| 724 |
+
"type": "text",
|
| 725 |
+
"text": "Then, soft masks $m _ { j , c }$ for the contribution of each example to each prototype are computed, by comparing to the threshold the normalized distances, as follows: ",
|
| 726 |
+
"bbox": [
|
| 727 |
+
174,
|
| 728 |
+
321,
|
| 729 |
+
823,
|
| 730 |
+
349
|
| 731 |
+
],
|
| 732 |
+
"page_idx": 5
|
| 733 |
+
},
|
| 734 |
+
{
|
| 735 |
+
"type": "equation",
|
| 736 |
+
"img_path": "images/38d3a56e18ad32868241f34ac545f48e2690001ff7d701ea90ef1ea781885cb8.jpg",
|
| 737 |
+
"text": "$$\n\\tilde { p } _ { c } = \\frac { \\sum _ { i } h ( { \\bf x } _ { i } ) z _ { i , c } + \\sum _ { j } h ( \\tilde { x } _ { j } ) \\tilde { z } _ { j , c } m _ { j , c } } { \\sum _ { i } z _ { i , c } + \\sum _ { j } \\tilde { z } _ { j , c } m _ { j , c } } , \\mathrm { ~ w h e r e ~ } m _ { j , c } = \\sigma \\left( - \\gamma _ { c } \\left( \\tilde { d } _ { j , c } - \\beta _ { c } \\right) \\right)\n$$",
|
| 738 |
+
"text_format": "latex",
|
| 739 |
+
"bbox": [
|
| 740 |
+
233,
|
| 741 |
+
353,
|
| 742 |
+
763,
|
| 743 |
+
393
|
| 744 |
+
],
|
| 745 |
+
"page_idx": 5
|
| 746 |
+
},
|
| 747 |
+
{
|
| 748 |
+
"type": "text",
|
| 749 |
+
"text": "where $\\sigma ( \\cdot )$ is the sigmoid function. ",
|
| 750 |
+
"bbox": [
|
| 751 |
+
174,
|
| 752 |
+
397,
|
| 753 |
+
405,
|
| 754 |
+
412
|
| 755 |
+
],
|
| 756 |
+
"page_idx": 5
|
| 757 |
+
},
|
| 758 |
+
{
|
| 759 |
+
"type": "text",
|
| 760 |
+
"text": "When training with this refinement process, the model can now use its MLP in Equation 8 to learn to include or ignore entirely certain unlabeled examples. The use of soft masks makes this process entirely differentiable2. Finally, much like for regular soft $k$ -means (with or without a distractor cluster), while we could recursively repeat the refinement for multiple steps, we found a single step to perform well enough. ",
|
| 761 |
+
"bbox": [
|
| 762 |
+
174,
|
| 763 |
+
419,
|
| 764 |
+
825,
|
| 765 |
+
489
|
| 766 |
+
],
|
| 767 |
+
"page_idx": 5
|
| 768 |
+
},
|
| 769 |
+
{
|
| 770 |
+
"type": "text",
|
| 771 |
+
"text": "4 RELATED WORK ",
|
| 772 |
+
"text_level": 1,
|
| 773 |
+
"bbox": [
|
| 774 |
+
176,
|
| 775 |
+
508,
|
| 776 |
+
344,
|
| 777 |
+
525
|
| 778 |
+
],
|
| 779 |
+
"page_idx": 5
|
| 780 |
+
},
|
| 781 |
+
{
|
| 782 |
+
"type": "text",
|
| 783 |
+
"text": "We summarize here the most relevant work from the literature on few-shot learning, semi-supervised learning and clustering. ",
|
| 784 |
+
"bbox": [
|
| 785 |
+
174,
|
| 786 |
+
540,
|
| 787 |
+
823,
|
| 788 |
+
569
|
| 789 |
+
],
|
| 790 |
+
"page_idx": 5
|
| 791 |
+
},
|
| 792 |
+
{
|
| 793 |
+
"type": "text",
|
| 794 |
+
"text": "The best performing methods for few-shot learning use the episodic training framework prescribed by meta-learning. The approach within which our work falls is that of metric learning methods. Previous work in metric-learning for few-shot-classification includes Deep Siamese Networks (Koch et al., 2015), Matching Networks (Vinyals et al., 2016), and Prototypical Networks (Snell et al., 2017), which is the model we extend to the semi-supervised setting in our work. The general idea here is to learn an embedding function that embeds examples belonging to the same class close together while keeping embeddings from separate classes far apart. Distances between embeddings of items from the support set and query set are then used as a notion of similarity to do classification. Lastly, closely related to our work with regard to extending the few-shot learning setting, Bachman et al. (2017) employ Matching Networks in an active learning framework where the model has a choice of which unlabeled item to add to the support set over a certain number of time steps before classifying the query set. Unlike our setting, their meta-learning agent can acquire ground-truth labels from the unlabeled set, and they do not use distractor examples. ",
|
| 795 |
+
"bbox": [
|
| 796 |
+
173,
|
| 797 |
+
575,
|
| 798 |
+
825,
|
| 799 |
+
756
|
| 800 |
+
],
|
| 801 |
+
"page_idx": 5
|
| 802 |
+
},
|
| 803 |
+
{
|
| 804 |
+
"type": "text",
|
| 805 |
+
"text": "Other meta-learning approaches to few-shot learning include learning how to use the support set to update a learner model so as to generalize to the query set. Recent work has involved learning either the weight initialization and/or update step that is used by a learner neural network (Ravi & Larochelle, 2017; Finn et al., 2017). Another approach is to train a generic neural architecture such as a memory-augmented recurrent network (Santoro et al., 2016) or a temporal convolutional network (Mishra et al., 2017) to sequentially process the support set and perform accurate predictions of the labels of the query set examples. These other methods are also competitive for few-shot learning, but we chose to extend Prototypical Networks in this work for its simplicity and efficiency. ",
|
| 806 |
+
"bbox": [
|
| 807 |
+
174,
|
| 808 |
+
762,
|
| 809 |
+
825,
|
| 810 |
+
888
|
| 811 |
+
],
|
| 812 |
+
"page_idx": 5
|
| 813 |
+
},
|
| 814 |
+
{
|
| 815 |
+
"type": "text",
|
| 816 |
+
"text": "As for the literature on semi-supervised learning, while it is quite vast (Zhu, 2005; Chapelle et al., 2010), the most relevant category to our work is related to self-training (Yarowsky, 1995; Rosenberg et al., 2005). Here, a classifier is first trained on the initial training set. The classifier is then used to classify unlabeled items, and the most confidently predicted unlabeled items are added to the training set with the prediction of the classifier as the assumed label. This is similar to our soft $k$ -Means extension to Prototypical Networks. Indeed, since the soft assignments (Equation 4) match the regular Prototypical Network’s classifier output for new inputs (Equation 2), then the refinement can be thought of re-feeding to a Prototypical Network a new support set augmented with (soft) self-labels from the unlabeled set. ",
|
| 817 |
+
"bbox": [
|
| 818 |
+
174,
|
| 819 |
+
103,
|
| 820 |
+
825,
|
| 821 |
+
228
|
| 822 |
+
],
|
| 823 |
+
"page_idx": 6
|
| 824 |
+
},
|
| 825 |
+
{
|
| 826 |
+
"type": "text",
|
| 827 |
+
"text": "Our algorithm is also related to transductive learning (Vapnik, 1998; Joachims, 1999; Fu et al., 2015), where the base classifier gets refined by seeing the unlabeled examples. In practice, one could use our method in a transductive setting where the unlabeled set is the same as the query set; however, here to avoid our model memorizing labels of the unlabeled set during the meta-learning procedure, we split out a separate unlabeled set that is different from the query set. ",
|
| 828 |
+
"bbox": [
|
| 829 |
+
174,
|
| 830 |
+
236,
|
| 831 |
+
823,
|
| 832 |
+
305
|
| 833 |
+
],
|
| 834 |
+
"page_idx": 6
|
| 835 |
+
},
|
| 836 |
+
{
|
| 837 |
+
"type": "text",
|
| 838 |
+
"text": "In addition to the original $k$ -Means method (Lloyd, 1982), the most related work to our setup involving clustering algorithms considers applying $k$ -Means in the presence of outliers (Hautamäki et al., 2005; Chawla & Gionis, 2013; Gupta et al., 2017). The goal here is to correctly discover and ignore the outliers so that they do not wrongly shift the cluster locations to form a bad partition of the true data. This objective is also important in our setup as not ignoring outliers (or distractors) will wrongly shift the prototypes and negatively influence classification performance. ",
|
| 839 |
+
"bbox": [
|
| 840 |
+
174,
|
| 841 |
+
311,
|
| 842 |
+
825,
|
| 843 |
+
396
|
| 844 |
+
],
|
| 845 |
+
"page_idx": 6
|
| 846 |
+
},
|
| 847 |
+
{
|
| 848 |
+
"type": "text",
|
| 849 |
+
"text": "Our contribution to the semi-supervised learning and clustering literature is to go beyond the classical setting of training and evaluating within a single dataset, and consider the setting where we must learn to transfer from a set of training classes $\\mathcal { C } _ { \\mathrm { t r a i n } }$ to a new set of test classes $\\mathcal { C } _ { \\mathrm { t e s t } }$ . ",
|
| 850 |
+
"bbox": [
|
| 851 |
+
176,
|
| 852 |
+
404,
|
| 853 |
+
823,
|
| 854 |
+
445
|
| 855 |
+
],
|
| 856 |
+
"page_idx": 6
|
| 857 |
+
},
|
| 858 |
+
{
|
| 859 |
+
"type": "text",
|
| 860 |
+
"text": "5 EXPERIMENTS ",
|
| 861 |
+
"text_level": 1,
|
| 862 |
+
"bbox": [
|
| 863 |
+
176,
|
| 864 |
+
467,
|
| 865 |
+
326,
|
| 866 |
+
483
|
| 867 |
+
],
|
| 868 |
+
"page_idx": 6
|
| 869 |
+
},
|
| 870 |
+
{
|
| 871 |
+
"type": "text",
|
| 872 |
+
"text": "5.1 DATASETS ",
|
| 873 |
+
"text_level": 1,
|
| 874 |
+
"bbox": [
|
| 875 |
+
174,
|
| 876 |
+
500,
|
| 877 |
+
287,
|
| 878 |
+
515
|
| 879 |
+
],
|
| 880 |
+
"page_idx": 6
|
| 881 |
+
},
|
| 882 |
+
{
|
| 883 |
+
"type": "text",
|
| 884 |
+
"text": "We evaluate the performance of our model on three datasets: two benchmark few-shot classification datasets and a novel large-scale dataset that we hope will be useful for future few-shot learning work. ",
|
| 885 |
+
"bbox": [
|
| 886 |
+
176,
|
| 887 |
+
526,
|
| 888 |
+
823,
|
| 889 |
+
555
|
| 890 |
+
],
|
| 891 |
+
"page_idx": 6
|
| 892 |
+
},
|
| 893 |
+
{
|
| 894 |
+
"type": "text",
|
| 895 |
+
"text": "Omniglot (Lake et al., 2011) is a dataset of 1,623 handwritten characters from 50 alphabets. Each character was drawn by 20 human subjects. We follow the few-shot setting proposed by Vinyals et al. (2016), in which the images are resized to $2 8 \\times 2 8$ pixels and rotations in multiples of $9 0 ^ { \\circ }$ are applied, yielding 6,492 classes in total. These are split into 4,112 training classes, 688 validation classes, and 1,692 testing classes. ",
|
| 896 |
+
"bbox": [
|
| 897 |
+
174,
|
| 898 |
+
561,
|
| 899 |
+
823,
|
| 900 |
+
631
|
| 901 |
+
],
|
| 902 |
+
"page_idx": 6
|
| 903 |
+
},
|
| 904 |
+
{
|
| 905 |
+
"type": "text",
|
| 906 |
+
"text": "miniImageNet (Vinyals et al., 2016) is a modified version of the ILSVRC-12 dataset (Russakovsky et al., 2015), in which 600 images for each of 100 classes were randomly chosen to be part of the dataset. We rely on the class split used by Ravi & Larochelle (2017). These splits use 64 classes for training, 16 for validation, and 20 for test. All images are of size $8 4 \\times 8 4$ pixels. ",
|
| 907 |
+
"bbox": [
|
| 908 |
+
174,
|
| 909 |
+
638,
|
| 910 |
+
823,
|
| 911 |
+
694
|
| 912 |
+
],
|
| 913 |
+
"page_idx": 6
|
| 914 |
+
},
|
| 915 |
+
{
|
| 916 |
+
"type": "text",
|
| 917 |
+
"text": "tieredImageNet is our proposed dataset for few-shot classification. Like miniImagenet, it is a subset of ILSVRC-12. However, tieredImageNet represents a larger subset of ILSVRC-12 (608 classes rather than 100 for miniImageNet). Analogous to Omniglot, in which characters are grouped into alphabets, tieredImageNet groups classes into broader categories corresponding to higher-level nodes in the ImageNet (Deng et al., 2009) hierarchy. There are 34 categories in total, with each category containing between 10 and 30 classes. These are split into 20 training, 6 validation and 8 testing categories (details of the dataset can be found in the supplementary material). This ensures that all of the training classes are sufficiently distinct from the testing classes, unlike miniImageNet and other alternatives such as randImageNet proposed by Vinyals et al. (2016). For example, “pipe organ” is a training class and “electric guitar” is a test class in the Ravi & Larochelle (2017) split of miniImagenet, even though they are both musical instruments. This scenario would not occur in tieredImageNet since “musical instrument” is a high-level category and as such is not split between training and test classes. This represents a more realistic few-shot learning scenario since in general we cannot assume that test classes will be similar to those seen in training. Additionally, the tiered structure of tieredImageNet may be useful for few-shot learning approaches that can take advantage of hierarchical relationships between classes. We leave such interesting extensions for future work. ",
|
| 918 |
+
"bbox": [
|
| 919 |
+
174,
|
| 920 |
+
702,
|
| 921 |
+
825,
|
| 922 |
+
922
|
| 923 |
+
],
|
| 924 |
+
"page_idx": 6
|
| 925 |
+
},
|
| 926 |
+
{
|
| 927 |
+
"type": "table",
|
| 928 |
+
"img_path": "images/f7b16d6629c0cbad8f28039697a9ca6f0afb364726c938546951a2bab885a499.jpg",
|
| 929 |
+
"table_caption": [
|
| 930 |
+
"Table 1: Omniglot 1-shot classification results. In this table as well as those below “w/ D” denotes “with distractors”, where the unlabeled images contain irrelevant classes. "
|
| 931 |
+
],
|
| 932 |
+
"table_footnote": [],
|
| 933 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Models</td><td rowspan=1 colspan=1>Acc.</td><td rowspan=1 colspan=1>Acc. w/ D</td></tr><tr><td rowspan=1 colspan=1>Supervised</td><td rowspan=1 colspan=1>94.62 ± 0.09</td><td rowspan=1 colspan=1>94.62 ± 0.09</td></tr><tr><td rowspan=1 colspan=1>Semi-Supervised Inference</td><td rowspan=1 colspan=1>97.45 ± 0.05</td><td rowspan=1 colspan=1>95.08 ± 0.09</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means</td><td rowspan=1 colspan=1>97.25 ± 0.10</td><td rowspan=1 colspan=1>95.01 ± 0.09</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means+Cluster</td><td rowspan=1 colspan=1>97.68 ± 0.07</td><td rowspan=1 colspan=1>97.17 ± 0.04</td></tr><tr><td rowspan=1 colspan=1>Masked Soft k-Means</td><td rowspan=1 colspan=1>97.52 ± 0.07</td><td rowspan=1 colspan=1>97.30 ± 0.08</td></tr></table>",
|
| 934 |
+
"bbox": [
|
| 935 |
+
292,
|
| 936 |
+
99,
|
| 937 |
+
705,
|
| 938 |
+
195
|
| 939 |
+
],
|
| 940 |
+
"page_idx": 7
|
| 941 |
+
},
|
| 942 |
+
{
|
| 943 |
+
"type": "text",
|
| 944 |
+
"text": "5.2 ADAPTING THE DATASETS FOR SEMI-SUPERVISED LEARNING",
|
| 945 |
+
"text_level": 1,
|
| 946 |
+
"bbox": [
|
| 947 |
+
176,
|
| 948 |
+
260,
|
| 949 |
+
642,
|
| 950 |
+
275
|
| 951 |
+
],
|
| 952 |
+
"page_idx": 7
|
| 953 |
+
},
|
| 954 |
+
{
|
| 955 |
+
"type": "text",
|
| 956 |
+
"text": "For each dataset, we first create an additional split to separate the images of each class into disjoint labeled and unlabeled sets. For Omniglot and tieredImageNet we sample $10 \\%$ of the images of each class to form the labeled split. The remaining $90 \\%$ can only be used in the unlabeled portion of episodes. For miniImageNet we use $40 \\%$ of the data for the labeled split and the remaining $60 \\%$ for the unlabeled, since we noticed that $10 \\%$ was too small to achieve reasonable performance and avoid overfitting. We report the average classification scores over 10 random splits of labeled and unlabeled portions of the training set, with uncertainty computed in standard error (standard deviation divided by the square root of the total number of splits). ",
|
| 957 |
+
"bbox": [
|
| 958 |
+
173,
|
| 959 |
+
286,
|
| 960 |
+
825,
|
| 961 |
+
398
|
| 962 |
+
],
|
| 963 |
+
"page_idx": 7
|
| 964 |
+
},
|
| 965 |
+
{
|
| 966 |
+
"type": "text",
|
| 967 |
+
"text": "We would like to emphasize that due to this labeled/unlabeled split, we are using strictly less label information than in the previously-published work on these datasets. Because of this, we do not expect our results to match the published numbers, which should instead be interpreted as an upperbound for the performance of the semi-supervised models defined in this work. ",
|
| 968 |
+
"bbox": [
|
| 969 |
+
174,
|
| 970 |
+
405,
|
| 971 |
+
825,
|
| 972 |
+
460
|
| 973 |
+
],
|
| 974 |
+
"page_idx": 7
|
| 975 |
+
},
|
| 976 |
+
{
|
| 977 |
+
"type": "text",
|
| 978 |
+
"text": "Episode construction then is performed as follows. For a given dataset, we create a training episode by first sampling $N$ classes uniformly at random from the set of training classes $\\mathcal { C } _ { \\mathrm { t r a i n } }$ . We then sample $K$ images from the labeled split of each of these classes to form the support set, and $M$ images from the unlabeled split of each of these classes to form the unlabeled set. Optionally, when including distractors, we additionally sample $H$ other classes from the set of training classes and $M$ images from the unlabeled split of each to act as the distractors. These distractor images are added to the unlabeled set along with the unlabeled images of the $N$ classes of interest (for a total of $M N + M H$ unlabeled images). The query portion of the episode is comprised of a fixed number of images from the labeled split of each of the $N$ chosen classes. Test episodes are created analogously, but with the $N$ classes (and optionally the $H$ distractor classes) sampled from $\\mathcal { C } _ { \\mathrm { t e s t } }$ . In the experiments reported here we used $H = N = 5$ , i.e. 5 classes for both the labeled classes and the distractor classes. We used $M = 5$ for training and $M = 2 0$ for testing in most cases, thus measuring the ability of the models to generalize to a larger unlabeled set size. Details of the dataset splits, including the specific classes assigned to train/validation/test sets, can be found in Appendices A and B. ",
|
| 979 |
+
"bbox": [
|
| 980 |
+
174,
|
| 981 |
+
468,
|
| 982 |
+
825,
|
| 983 |
+
675
|
| 984 |
+
],
|
| 985 |
+
"page_idx": 7
|
| 986 |
+
},
|
| 987 |
+
{
|
| 988 |
+
"type": "text",
|
| 989 |
+
"text": "In each dataset we compare our three semi-supervised models with two baselines. The first baseline, referred to as “Supervised” in our tables, is an ordinary Prototypical Network that is trained in a purely supervised way on the labeled split of each dataset. The second baseline, referred to as “Semi-Supervised Inference”, uses the embedding function learned by this supervised Prototypical Network, but performs semi-supervised refinement of the prototypes at test time using a step of Soft $k$ -Means refinement. This is to be contrasted with our semi-supervised models that perform this refinement both at training time and at test time, therefore learning a different embedding function. We evaluate each model in two settings: one where all unlabeled examples belong to the classes of interest, and a more challenging one that includes distractors. Details of the model hyperparameters can be found in Appendix D and our online repository.3 ",
|
| 990 |
+
"bbox": [
|
| 991 |
+
174,
|
| 992 |
+
683,
|
| 993 |
+
825,
|
| 994 |
+
821
|
| 995 |
+
],
|
| 996 |
+
"page_idx": 7
|
| 997 |
+
},
|
| 998 |
+
{
|
| 999 |
+
"type": "text",
|
| 1000 |
+
"text": "5.3 RESULTS ",
|
| 1001 |
+
"text_level": 1,
|
| 1002 |
+
"bbox": [
|
| 1003 |
+
174,
|
| 1004 |
+
840,
|
| 1005 |
+
277,
|
| 1006 |
+
854
|
| 1007 |
+
],
|
| 1008 |
+
"page_idx": 7
|
| 1009 |
+
},
|
| 1010 |
+
{
|
| 1011 |
+
"type": "text",
|
| 1012 |
+
"text": "Results for Omniglot, miniImageNet and tieredImageNet are given in Tables 1, 2 and 5, respectively, while Figure 4 shows the performance of our models on tieredImageNet (our largest dataset) using different values for $M$ (number of items in the unlabeled set per class). Additional results comparing the ProtoNet model to various baselines on these datasets, and analysis of the performance of the Masked Soft $k$ -Means model can be found in Appendix C. ",
|
| 1013 |
+
"bbox": [
|
| 1014 |
+
176,
|
| 1015 |
+
867,
|
| 1016 |
+
823,
|
| 1017 |
+
896
|
| 1018 |
+
],
|
| 1019 |
+
"page_idx": 7
|
| 1020 |
+
},
|
| 1021 |
+
{
|
| 1022 |
+
"type": "table",
|
| 1023 |
+
"img_path": "images/31ac5a29eac847dd2cbac3f4f93c3a7df23712f9f64b093aae1f167b2cfa0e56.jpg",
|
| 1024 |
+
"table_caption": [
|
| 1025 |
+
"Table 2: miniImageNet 1/5-shot classification results. "
|
| 1026 |
+
],
|
| 1027 |
+
"table_footnote": [],
|
| 1028 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Models</td><td rowspan=1 colspan=1>1-shot Acc.</td><td rowspan=1 colspan=1>5-shot Acc.</td><td rowspan=1 colspan=1>1-shot Acc w/ D</td><td rowspan=1 colspan=1> 5-shot Acc. w/ D</td></tr><tr><td rowspan=1 colspan=1>Supervised</td><td rowspan=1 colspan=1>43.61 ± 0.27</td><td rowspan=1 colspan=1>59.08 ± 0.22</td><td rowspan=1 colspan=1>43.61 ± 0.27</td><td rowspan=1 colspan=1>59.08 ± 0.22</td></tr><tr><td rowspan=1 colspan=1>Semi-Supervised Inference</td><td rowspan=1 colspan=1>48.98 ± 0.34</td><td rowspan=1 colspan=1>63.77 ± 0.20</td><td rowspan=1 colspan=1>47.42 ± 0.33</td><td rowspan=1 colspan=1>62.62 ± 0.24</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means</td><td rowspan=1 colspan=1>50.09 ± 0.45</td><td rowspan=1 colspan=1>64.59 ± 0.28</td><td rowspan=1 colspan=1>48.70 ± 0.32</td><td rowspan=1 colspan=1>63.55 ± 0.28</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means+Cluster</td><td rowspan=1 colspan=1>49.03 ± 0.24</td><td rowspan=1 colspan=1>63.08 ± 0.18</td><td rowspan=1 colspan=1>48.86± 0.32</td><td rowspan=1 colspan=1>61.27 ± 0.24</td></tr><tr><td rowspan=1 colspan=1>Masked Soft k-Means</td><td rowspan=1 colspan=1>50.41 ± 0.31</td><td rowspan=1 colspan=1>64.39 ± 0.24</td><td rowspan=1 colspan=1>49.04 ± 0.31</td><td rowspan=1 colspan=1>62.96 ± 0.14</td></tr></table>",
|
| 1029 |
+
"bbox": [
|
| 1030 |
+
176,
|
| 1031 |
+
99,
|
| 1032 |
+
820,
|
| 1033 |
+
191
|
| 1034 |
+
],
|
| 1035 |
+
"page_idx": 8
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"type": "table",
|
| 1039 |
+
"img_path": "images/ad57873476ec47ba8aedc676591637cf226e3c936d80d20b1ce99d40edd18e4d.jpg",
|
| 1040 |
+
"table_caption": [
|
| 1041 |
+
"Table 3: tieredImageNet 1/5-shot classification results. "
|
| 1042 |
+
],
|
| 1043 |
+
"table_footnote": [],
|
| 1044 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Models</td><td rowspan=1 colspan=1>1-shot Acc.</td><td rowspan=1 colspan=1> 5-shot Acc.</td><td rowspan=1 colspan=1>1-shot Acc. w/ D</td><td rowspan=1 colspan=1> 5-shot Acc. w/ D</td></tr><tr><td rowspan=1 colspan=1>Supervised</td><td rowspan=1 colspan=1>46.52 ± 0.52</td><td rowspan=1 colspan=1>66.15 ± 0.22</td><td rowspan=1 colspan=1>46.52 ± 0.52</td><td rowspan=1 colspan=1>66.15 ± 0.22</td></tr><tr><td rowspan=1 colspan=1>Semi-Supervised Inference</td><td rowspan=1 colspan=1>50.74 ± 0.75</td><td rowspan=1 colspan=1>69.37 ± 0.26</td><td rowspan=1 colspan=1>48.67 ± 0.60</td><td rowspan=1 colspan=1>67.46 ± 0.24</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means</td><td rowspan=1 colspan=1>51.52 ± 0.36</td><td rowspan=1 colspan=1>70.25 ± 0.31</td><td rowspan=1 colspan=1>49.88 ± 0.52</td><td rowspan=1 colspan=1>68.32 ± 0.22</td></tr><tr><td rowspan=1 colspan=1>Soft k-Means+Cluster</td><td rowspan=1 colspan=1>51.85 ± 0.25</td><td rowspan=1 colspan=1>69.42 ± 0.17</td><td rowspan=1 colspan=1>51.36 ± 0.31</td><td rowspan=1 colspan=1>67.56 ± 0.10</td></tr><tr><td rowspan=1 colspan=1>Masked Soft k-Means</td><td rowspan=1 colspan=1>52.39 ± 0.44</td><td rowspan=1 colspan=1>69.88 ± 0.20</td><td rowspan=1 colspan=1>51.38 ± 0.38</td><td rowspan=1 colspan=1>69.08 ± 0.25</td></tr></table>",
|
| 1045 |
+
"bbox": [
|
| 1046 |
+
178,
|
| 1047 |
+
236,
|
| 1048 |
+
821,
|
| 1049 |
+
327
|
| 1050 |
+
],
|
| 1051 |
+
"page_idx": 8
|
| 1052 |
+
},
|
| 1053 |
+
{
|
| 1054 |
+
"type": "text",
|
| 1055 |
+
"text": "",
|
| 1056 |
+
"bbox": [
|
| 1057 |
+
174,
|
| 1058 |
+
386,
|
| 1059 |
+
825,
|
| 1060 |
+
428
|
| 1061 |
+
],
|
| 1062 |
+
"page_idx": 8
|
| 1063 |
+
},
|
| 1064 |
+
{
|
| 1065 |
+
"type": "text",
|
| 1066 |
+
"text": "Across all three benchmarks, at least one of our proposed models outperforms the baselines, demonstrating the effectiveness of our semi-supervised meta-learning procedure. In the nondistractor settings, all three proposed models outperform the baselines in almost all the experiments, without a clear winner between the three models across the datasets and shot numbers. In the scenario where training and testing includes distractors, Masked Soft $k$ -Means shows the most robust performance across all three datasets, attaining the best results in each case but one. In fact this model reaches performance that is close to the upper bound based on the results without distractors. ",
|
| 1067 |
+
"bbox": [
|
| 1068 |
+
174,
|
| 1069 |
+
435,
|
| 1070 |
+
825,
|
| 1071 |
+
546
|
| 1072 |
+
],
|
| 1073 |
+
"page_idx": 8
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"type": "text",
|
| 1077 |
+
"text": "From Figure 4, we observe clear improvements in test accuracy when the number of items in the unlabeled set per class grows from 0 to 25. These models were trained with $M = 5$ and thus are showing an ability to extrapolate in generalization. This confirms that, through meta-training, the models learn to acquire a better representation that is improved by semi-supervised refinement. ",
|
| 1078 |
+
"bbox": [
|
| 1079 |
+
174,
|
| 1080 |
+
554,
|
| 1081 |
+
825,
|
| 1082 |
+
609
|
| 1083 |
+
],
|
| 1084 |
+
"page_idx": 8
|
| 1085 |
+
},
|
| 1086 |
+
{
|
| 1087 |
+
"type": "text",
|
| 1088 |
+
"text": "6 CONCLUSION ",
|
| 1089 |
+
"text_level": 1,
|
| 1090 |
+
"bbox": [
|
| 1091 |
+
174,
|
| 1092 |
+
638,
|
| 1093 |
+
318,
|
| 1094 |
+
655
|
| 1095 |
+
],
|
| 1096 |
+
"page_idx": 8
|
| 1097 |
+
},
|
| 1098 |
+
{
|
| 1099 |
+
"type": "text",
|
| 1100 |
+
"text": "In this work, we propose a novel semi-supervised few-shot learning paradigm, where an unlabeled set is added to each episode. We also extend the setup to more realistic situations where the unlabeled set has novel classes distinct from the labeled classes. To address the problem that current fewshot classification datasets are too small for a labeled vs. unlabeled split and also lack hierarchical levels of labels, we introduce a new dataset, tieredImageNet. We propose several novel extensions of Prototypical Networks, and they show consistent improvements under semi-supervised settings compared to our baselines. As future work, we are working on incorporating fast weights (Ba et al., 2016; Finn et al., 2017) into our framework so that examples can have different embedding representations given the contents in the episode. ",
|
| 1101 |
+
"bbox": [
|
| 1102 |
+
174,
|
| 1103 |
+
676,
|
| 1104 |
+
825,
|
| 1105 |
+
801
|
| 1106 |
+
],
|
| 1107 |
+
"page_idx": 8
|
| 1108 |
+
},
|
| 1109 |
+
{
|
| 1110 |
+
"type": "text",
|
| 1111 |
+
"text": "Acknowledgement Supported by grants from NSERC, Samsung, and the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/Interior Business Center (DoI/IBC) contract number D16PC00003. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DoI/IBC, or the U.S. Government. ",
|
| 1112 |
+
"bbox": [
|
| 1113 |
+
174,
|
| 1114 |
+
825,
|
| 1115 |
+
825,
|
| 1116 |
+
922
|
| 1117 |
+
],
|
| 1118 |
+
"page_idx": 8
|
| 1119 |
+
},
|
| 1120 |
+
{
|
| 1121 |
+
"type": "image",
|
| 1122 |
+
"img_path": "images/efb36bc0731609f9e536fc139ba587b8e6c9de34b7909d4c66be59f685b76cab.jpg",
|
| 1123 |
+
"image_caption": [
|
| 1124 |
+
"Figure 4: Model Performance on tieredImageNet with different numbers of unlabeled items during test time. "
|
| 1125 |
+
],
|
| 1126 |
+
"image_footnote": [],
|
| 1127 |
+
"bbox": [
|
| 1128 |
+
176,
|
| 1129 |
+
104,
|
| 1130 |
+
820,
|
| 1131 |
+
349
|
| 1132 |
+
],
|
| 1133 |
+
"page_idx": 9
|
| 1134 |
+
},
|
| 1135 |
+
{
|
| 1136 |
+
"type": "text",
|
| 1137 |
+
"text": "REFERENCES ",
|
| 1138 |
+
"text_level": 1,
|
| 1139 |
+
"bbox": [
|
| 1140 |
+
174,
|
| 1141 |
+
404,
|
| 1142 |
+
285,
|
| 1143 |
+
419
|
| 1144 |
+
],
|
| 1145 |
+
"page_idx": 9
|
| 1146 |
+
},
|
| 1147 |
+
{
|
| 1148 |
+
"type": "text",
|
| 1149 |
+
"text": "Jimmy Ba, Geoffrey E. Hinton, Volodymyr Mnih, Joel Z. Leibo, and Catalin Ionescu. Using fast weights to attend to the recent past. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pp. 4331–4339, 2016. ",
|
| 1150 |
+
"bbox": [
|
| 1151 |
+
174,
|
| 1152 |
+
428,
|
| 1153 |
+
825,
|
| 1154 |
+
484
|
| 1155 |
+
],
|
| 1156 |
+
"page_idx": 9
|
| 1157 |
+
},
|
| 1158 |
+
{
|
| 1159 |
+
"type": "text",
|
| 1160 |
+
"text": "Philip Bachman, Alessandro Sordoni, and Adam Trischler. Learning algorithms for active learning. 2017. ",
|
| 1161 |
+
"bbox": [
|
| 1162 |
+
173,
|
| 1163 |
+
494,
|
| 1164 |
+
821,
|
| 1165 |
+
523
|
| 1166 |
+
],
|
| 1167 |
+
"page_idx": 9
|
| 1168 |
+
},
|
| 1169 |
+
{
|
| 1170 |
+
"type": "text",
|
| 1171 |
+
"text": "Olivier Chapelle, Bernhard Schölkopf, and Alexander Zien. Semi-Supervised Learning. The MIT Press, 1st edition, 2010. ISBN 0262514125, 9780262514125. ",
|
| 1172 |
+
"bbox": [
|
| 1173 |
+
174,
|
| 1174 |
+
535,
|
| 1175 |
+
823,
|
| 1176 |
+
564
|
| 1177 |
+
],
|
| 1178 |
+
"page_idx": 9
|
| 1179 |
+
},
|
| 1180 |
+
{
|
| 1181 |
+
"type": "text",
|
| 1182 |
+
"text": "Sanjay Chawla and Aristides Gionis. k-means–: A unified approach to clustering and outlier detection. In Proceedings of the 2013 SIAM International Conference on Data Mining, pp. 189– 197. SIAM, 2013. ",
|
| 1183 |
+
"bbox": [
|
| 1184 |
+
173,
|
| 1185 |
+
574,
|
| 1186 |
+
823,
|
| 1187 |
+
617
|
| 1188 |
+
],
|
| 1189 |
+
"page_idx": 9
|
| 1190 |
+
},
|
| 1191 |
+
{
|
| 1192 |
+
"type": "text",
|
| 1193 |
+
"text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pp. 248–255. IEEE, 2009. ",
|
| 1194 |
+
"bbox": [
|
| 1195 |
+
173,
|
| 1196 |
+
627,
|
| 1197 |
+
823,
|
| 1198 |
+
671
|
| 1199 |
+
],
|
| 1200 |
+
"page_idx": 9
|
| 1201 |
+
},
|
| 1202 |
+
{
|
| 1203 |
+
"type": "text",
|
| 1204 |
+
"text": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In 34th International Conference on Machine Learning, 2017. ",
|
| 1205 |
+
"bbox": [
|
| 1206 |
+
173,
|
| 1207 |
+
681,
|
| 1208 |
+
821,
|
| 1209 |
+
712
|
| 1210 |
+
],
|
| 1211 |
+
"page_idx": 9
|
| 1212 |
+
},
|
| 1213 |
+
{
|
| 1214 |
+
"type": "text",
|
| 1215 |
+
"text": "Yanwei Fu, Timothy M. Hospedales, Tao Xiang, and Shaogang Gong. Transductive multi-view zero-shot learning. IEEE Trans. Pattern Anal. Mach. Intell., 37(11):2332–2345, 2015. ",
|
| 1216 |
+
"bbox": [
|
| 1217 |
+
174,
|
| 1218 |
+
722,
|
| 1219 |
+
823,
|
| 1220 |
+
751
|
| 1221 |
+
],
|
| 1222 |
+
"page_idx": 9
|
| 1223 |
+
},
|
| 1224 |
+
{
|
| 1225 |
+
"type": "text",
|
| 1226 |
+
"text": "Shalmoli Gupta, Ravi Kumar, Kefu Lu, Benjamin Moseley, and Sergei Vassilvitskii. Local search methods for k-means with outliers. Proceedings of the VLDB Endowment, 10(7):757–768, 2017. ",
|
| 1227 |
+
"bbox": [
|
| 1228 |
+
174,
|
| 1229 |
+
762,
|
| 1230 |
+
823,
|
| 1231 |
+
791
|
| 1232 |
+
],
|
| 1233 |
+
"page_idx": 9
|
| 1234 |
+
},
|
| 1235 |
+
{
|
| 1236 |
+
"type": "text",
|
| 1237 |
+
"text": "Ville Hautamäki, Svetlana Cherednichenko, Ismo Kärkkäinen, Tomi Kinnunen, and Pasi Fränti. Improving k-means by outlier removal. In Scandinavian Conference on Image Analysis, pp. 978– 987. Springer, 2005. ",
|
| 1238 |
+
"bbox": [
|
| 1239 |
+
174,
|
| 1240 |
+
801,
|
| 1241 |
+
823,
|
| 1242 |
+
844
|
| 1243 |
+
],
|
| 1244 |
+
"page_idx": 9
|
| 1245 |
+
},
|
| 1246 |
+
{
|
| 1247 |
+
"type": "text",
|
| 1248 |
+
"text": "Sepp Hochreiter, A Steven Younger, and Peter R Conwell. Learning to learn using gradient descent. In International Conference on Artificial Neural Networks, pp. 87–94. Springer, 2001. ",
|
| 1249 |
+
"bbox": [
|
| 1250 |
+
173,
|
| 1251 |
+
854,
|
| 1252 |
+
823,
|
| 1253 |
+
885
|
| 1254 |
+
],
|
| 1255 |
+
"page_idx": 9
|
| 1256 |
+
},
|
| 1257 |
+
{
|
| 1258 |
+
"type": "text",
|
| 1259 |
+
"text": "Thorsten Joachims. Transductive inference for text classification using support vector machines. In Proceedings of the Sixteenth International Conference on Machine Learning, 1999. ",
|
| 1260 |
+
"bbox": [
|
| 1261 |
+
176,
|
| 1262 |
+
895,
|
| 1263 |
+
823,
|
| 1264 |
+
924
|
| 1265 |
+
],
|
| 1266 |
+
"page_idx": 9
|
| 1267 |
+
},
|
| 1268 |
+
{
|
| 1269 |
+
"type": "text",
|
| 1270 |
+
"text": "Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. ",
|
| 1271 |
+
"bbox": [
|
| 1272 |
+
171,
|
| 1273 |
+
103,
|
| 1274 |
+
825,
|
| 1275 |
+
132
|
| 1276 |
+
],
|
| 1277 |
+
"page_idx": 10
|
| 1278 |
+
},
|
| 1279 |
+
{
|
| 1280 |
+
"type": "text",
|
| 1281 |
+
"text": "Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. Siamese neural networks for one-shot image recognition. In ICML Deep Learning Workshop, volume 2, 2015. ",
|
| 1282 |
+
"bbox": [
|
| 1283 |
+
171,
|
| 1284 |
+
140,
|
| 1285 |
+
823,
|
| 1286 |
+
170
|
| 1287 |
+
],
|
| 1288 |
+
"page_idx": 10
|
| 1289 |
+
},
|
| 1290 |
+
{
|
| 1291 |
+
"type": "text",
|
| 1292 |
+
"text": "Brenden M. Lake, Ruslan Salakhutdinov, Jason Gross, and Joshua B. Tenenbaum. One shot learning of simple visual concepts. In Proceedings of the 33th Annual Meeting of the Cognitive Science Society, CogSci 2011, Boston, Massachusetts, USA, July 20-23, 2011, 2011. ",
|
| 1293 |
+
"bbox": [
|
| 1294 |
+
176,
|
| 1295 |
+
178,
|
| 1296 |
+
821,
|
| 1297 |
+
222
|
| 1298 |
+
],
|
| 1299 |
+
"page_idx": 10
|
| 1300 |
+
},
|
| 1301 |
+
{
|
| 1302 |
+
"type": "text",
|
| 1303 |
+
"text": "Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory, 28(2): 129–137, 1982. ",
|
| 1304 |
+
"bbox": [
|
| 1305 |
+
173,
|
| 1306 |
+
229,
|
| 1307 |
+
823,
|
| 1308 |
+
258
|
| 1309 |
+
],
|
| 1310 |
+
"page_idx": 10
|
| 1311 |
+
},
|
| 1312 |
+
{
|
| 1313 |
+
"type": "text",
|
| 1314 |
+
"text": "Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. Meta-learning with temporal convolutions. CoRR, abs/1707.03141, 2017. ",
|
| 1315 |
+
"bbox": [
|
| 1316 |
+
169,
|
| 1317 |
+
267,
|
| 1318 |
+
823,
|
| 1319 |
+
297
|
| 1320 |
+
],
|
| 1321 |
+
"page_idx": 10
|
| 1322 |
+
},
|
| 1323 |
+
{
|
| 1324 |
+
"type": "text",
|
| 1325 |
+
"text": "Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In 5th International Conference on Learning Representations, 2017. ",
|
| 1326 |
+
"bbox": [
|
| 1327 |
+
171,
|
| 1328 |
+
305,
|
| 1329 |
+
823,
|
| 1330 |
+
335
|
| 1331 |
+
],
|
| 1332 |
+
"page_idx": 10
|
| 1333 |
+
},
|
| 1334 |
+
{
|
| 1335 |
+
"type": "text",
|
| 1336 |
+
"text": "Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. 2005. ",
|
| 1337 |
+
"bbox": [
|
| 1338 |
+
173,
|
| 1339 |
+
343,
|
| 1340 |
+
823,
|
| 1341 |
+
372
|
| 1342 |
+
],
|
| 1343 |
+
"page_idx": 10
|
| 1344 |
+
},
|
| 1345 |
+
{
|
| 1346 |
+
"type": "text",
|
| 1347 |
+
"text": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015. ",
|
| 1348 |
+
"bbox": [
|
| 1349 |
+
176,
|
| 1350 |
+
381,
|
| 1351 |
+
821,
|
| 1352 |
+
425
|
| 1353 |
+
],
|
| 1354 |
+
"page_idx": 10
|
| 1355 |
+
},
|
| 1356 |
+
{
|
| 1357 |
+
"type": "text",
|
| 1358 |
+
"text": "Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy P. Lillicrap. Oneshot learning with memory-augmented neural networks. In 33rd International Conference on Machine Learning, 2016. ",
|
| 1359 |
+
"bbox": [
|
| 1360 |
+
174,
|
| 1361 |
+
433,
|
| 1362 |
+
821,
|
| 1363 |
+
476
|
| 1364 |
+
],
|
| 1365 |
+
"page_idx": 10
|
| 1366 |
+
},
|
| 1367 |
+
{
|
| 1368 |
+
"type": "text",
|
| 1369 |
+
"text": "Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems 30, 2017. ",
|
| 1370 |
+
"bbox": [
|
| 1371 |
+
173,
|
| 1372 |
+
484,
|
| 1373 |
+
821,
|
| 1374 |
+
513
|
| 1375 |
+
],
|
| 1376 |
+
"page_idx": 10
|
| 1377 |
+
},
|
| 1378 |
+
{
|
| 1379 |
+
"type": "text",
|
| 1380 |
+
"text": "Sebastian Thrun. Lifelong learning algorithms. In Learning to learn, pp. 181–209. Springer, 1998. ",
|
| 1381 |
+
"bbox": [
|
| 1382 |
+
171,
|
| 1383 |
+
522,
|
| 1384 |
+
818,
|
| 1385 |
+
537
|
| 1386 |
+
],
|
| 1387 |
+
"page_idx": 10
|
| 1388 |
+
},
|
| 1389 |
+
{
|
| 1390 |
+
"type": "text",
|
| 1391 |
+
"text": "V.N. Vapnik. Statistical Learning Theory. Wiley, 1998. ISBN 9788126528929. ",
|
| 1392 |
+
"bbox": [
|
| 1393 |
+
174,
|
| 1394 |
+
545,
|
| 1395 |
+
699,
|
| 1396 |
+
561
|
| 1397 |
+
],
|
| 1398 |
+
"page_idx": 10
|
| 1399 |
+
},
|
| 1400 |
+
{
|
| 1401 |
+
"type": "text",
|
| 1402 |
+
"text": "Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems 29, pp. 3630–3638, 2016. ",
|
| 1403 |
+
"bbox": [
|
| 1404 |
+
173,
|
| 1405 |
+
569,
|
| 1406 |
+
823,
|
| 1407 |
+
613
|
| 1408 |
+
],
|
| 1409 |
+
"page_idx": 10
|
| 1410 |
+
},
|
| 1411 |
+
{
|
| 1412 |
+
"type": "text",
|
| 1413 |
+
"text": "David Yarowsky. Unsupervised word sense disambiguation rivaling supervised methods. In Proceedings of the 33rd annual meeting on Association for Computational Linguistics, pp. 189– 196. Association for Computational Linguistics, 1995. ",
|
| 1414 |
+
"bbox": [
|
| 1415 |
+
173,
|
| 1416 |
+
621,
|
| 1417 |
+
826,
|
| 1418 |
+
665
|
| 1419 |
+
],
|
| 1420 |
+
"page_idx": 10
|
| 1421 |
+
},
|
| 1422 |
+
{
|
| 1423 |
+
"type": "text",
|
| 1424 |
+
"text": "Xiaojin Zhu. Semi-supervised learning literature survey. 2005. ",
|
| 1425 |
+
"bbox": [
|
| 1426 |
+
173,
|
| 1427 |
+
674,
|
| 1428 |
+
588,
|
| 1429 |
+
689
|
| 1430 |
+
],
|
| 1431 |
+
"page_idx": 10
|
| 1432 |
+
},
|
| 1433 |
+
{
|
| 1434 |
+
"type": "text",
|
| 1435 |
+
"text": "A OMNIGLOT DATASET DETAILS ",
|
| 1436 |
+
"text_level": 1,
|
| 1437 |
+
"bbox": [
|
| 1438 |
+
176,
|
| 1439 |
+
102,
|
| 1440 |
+
464,
|
| 1441 |
+
118
|
| 1442 |
+
],
|
| 1443 |
+
"page_idx": 11
|
| 1444 |
+
},
|
| 1445 |
+
{
|
| 1446 |
+
"type": "text",
|
| 1447 |
+
"text": "We used the following split details for experiments on Omniglot dataset. This is the same train/test split as (Vinyals et al., 2016), but we created our own validation split for selecting hyper-parameters. Models are trained on the train split only. ",
|
| 1448 |
+
"bbox": [
|
| 1449 |
+
174,
|
| 1450 |
+
145,
|
| 1451 |
+
825,
|
| 1452 |
+
186
|
| 1453 |
+
],
|
| 1454 |
+
"page_idx": 11
|
| 1455 |
+
},
|
| 1456 |
+
{
|
| 1457 |
+
"type": "text",
|
| 1458 |
+
"text": "Train Alphabets: Alphabet_of_the_Magi, Angelic, Anglo-Saxon_Futhorc, Arcadian, Asomtavruli_(Georgian), Atemayar_Qelisayer, Atlantean, Aurek-Besh, Avesta, Balinese, Blackfoot_(Canadian_Aboriginal_Syllabics), Braille, Burmese_(Myanmar), Cyrillic, Futurama, Ge_ez, Glagolitic, Grantha, Greek, Gujarati, Gurmukhi (character 01-41), Inuktitut_(Canadian_Aboriginal_Syllabics), Japanese_(hiragana), Japanese_(katakana), Korean, Latin, Malay_(Jawi_-_Arabic), N_Ko, Ojibwe_(Canadian_Aboriginal_Syllabics), Sanskrit, Syriac_(Estrangelo), Tagalog, Tifinagh ",
|
| 1459 |
+
"bbox": [
|
| 1460 |
+
173,
|
| 1461 |
+
194,
|
| 1462 |
+
826,
|
| 1463 |
+
292
|
| 1464 |
+
],
|
| 1465 |
+
"page_idx": 11
|
| 1466 |
+
},
|
| 1467 |
+
{
|
| 1468 |
+
"type": "text",
|
| 1469 |
+
"text": "Validation Alphabets: Armenian, Bengali, Early_Aramaic, Hebrew, Mkhedruli_(Geogian) ",
|
| 1470 |
+
"bbox": [
|
| 1471 |
+
171,
|
| 1472 |
+
297,
|
| 1473 |
+
776,
|
| 1474 |
+
313
|
| 1475 |
+
],
|
| 1476 |
+
"page_idx": 11
|
| 1477 |
+
},
|
| 1478 |
+
{
|
| 1479 |
+
"type": "text",
|
| 1480 |
+
"text": "Test Alphabets: Gurmukhi (character 42-45), Kannada, Keble, Malayalam, Manipuri, Mongolian, Old_Church_Slavonic_(Cyrillic), Oriya, Sylheti, Syriac_(Serto), Tengwar, Tibetan, ULOG ",
|
| 1481 |
+
"bbox": [
|
| 1482 |
+
174,
|
| 1483 |
+
319,
|
| 1484 |
+
821,
|
| 1485 |
+
348
|
| 1486 |
+
],
|
| 1487 |
+
"page_idx": 11
|
| 1488 |
+
},
|
| 1489 |
+
{
|
| 1490 |
+
"type": "text",
|
| 1491 |
+
"text": "B tieredIMAGENET DATASET DETAILS ",
|
| 1492 |
+
"text_level": 1,
|
| 1493 |
+
"bbox": [
|
| 1494 |
+
174,
|
| 1495 |
+
387,
|
| 1496 |
+
504,
|
| 1497 |
+
404
|
| 1498 |
+
],
|
| 1499 |
+
"page_idx": 11
|
| 1500 |
+
},
|
| 1501 |
+
{
|
| 1502 |
+
"type": "text",
|
| 1503 |
+
"text": "Each high-level category in tieredImageNet contains between 10 and 30 ILSVRC-12 classes (17.8 on average). In the ImageNet hierarchy, some classes have multiple parent nodes. Therefore, classes belonging to more than one category were removed from the dataset to ensure separation between training and test categories. Test categories were chosen to reflect various levels of separation between training and test classes. Some test categories (such as “working dog”) are fairly similar to training categories, whereas others (such as “geological formation”) are quite different. The list of categories is shown below and statistics of the dataset can be found in Table 4. A visualization of the categories according to the ImageNet hierarchy is shown in Figure 5. The full list of classes per category will also be made public, however for the sake of brevity we do not include it here. ",
|
| 1504 |
+
"bbox": [
|
| 1505 |
+
173,
|
| 1506 |
+
430,
|
| 1507 |
+
825,
|
| 1508 |
+
556
|
| 1509 |
+
],
|
| 1510 |
+
"page_idx": 11
|
| 1511 |
+
},
|
| 1512 |
+
{
|
| 1513 |
+
"type": "table",
|
| 1514 |
+
"img_path": "images/a964613a6403b94754b999625e713b6a578e79d806b7164b2b4f183c32a0a494.jpg",
|
| 1515 |
+
"table_caption": [
|
| 1516 |
+
"Table 4: Statistics of the tieredImageNet dataset. "
|
| 1517 |
+
],
|
| 1518 |
+
"table_footnote": [],
|
| 1519 |
+
"table_body": "<table><tr><td></td><td>Train</td><td>Val</td><td>Test</td><td>Total</td></tr><tr><td>Categories</td><td>20</td><td>6</td><td>8</td><td>34</td></tr><tr><td>Classes</td><td>351</td><td>97</td><td>160</td><td>608</td></tr><tr><td>Images</td><td>448,695</td><td>124,261</td><td>206,209</td><td>779,165</td></tr></table>",
|
| 1520 |
+
"bbox": [
|
| 1521 |
+
308,
|
| 1522 |
+
609,
|
| 1523 |
+
689,
|
| 1524 |
+
670
|
| 1525 |
+
],
|
| 1526 |
+
"page_idx": 11
|
| 1527 |
+
},
|
| 1528 |
+
{
|
| 1529 |
+
"type": "text",
|
| 1530 |
+
"text": "Train Categories: n02087551 (hound, hound dog), n02092468 (terrier), n02120997 (feline, felid), n02370806 (ungulate, hoofed mammal), n02469914 (primate), n01726692 (snake, serpent, ophidian), n01674216 (saurian), n01524359 (passerine, passeriform bird), n01844917 (aquatic bird), n04081844 (restraint, constraint), n03574816 (instrument), n03800933 (musical instrument, instrument), n03125870 (craft), n04451818 (tool), n03414162 (game equipment), n03278248 (electronic equipment), n03419014 (garment), n03297735 (establishment), n02913152 (building, edifice), n04014297 (protective covering, protective cover, protection). ",
|
| 1531 |
+
"bbox": [
|
| 1532 |
+
173,
|
| 1533 |
+
700,
|
| 1534 |
+
825,
|
| 1535 |
+
813
|
| 1536 |
+
],
|
| 1537 |
+
"page_idx": 11
|
| 1538 |
+
},
|
| 1539 |
+
{
|
| 1540 |
+
"type": "text",
|
| 1541 |
+
"text": "Validation Categories: n02098550 (sporting dog, gun dog), n03257877 (durables, durable goods, consumer durables), n03405265 (furnishing), n03699975 (machine), n03738472 (mechanism), n03791235 (motor vehicle, automotive vehicle), ",
|
| 1542 |
+
"bbox": [
|
| 1543 |
+
174,
|
| 1544 |
+
819,
|
| 1545 |
+
825,
|
| 1546 |
+
861
|
| 1547 |
+
],
|
| 1548 |
+
"page_idx": 11
|
| 1549 |
+
},
|
| 1550 |
+
{
|
| 1551 |
+
"type": "text",
|
| 1552 |
+
"text": "Test Categories: n02103406 (working dog), n01473806 (aquatic vertebrate), n02159955 (insect), n04531098 (vessel), n03839993 (obstruction, obstructor, obstructer, impediment, impedimenta), n09287968 (geological formation, formation), n00020090 (substance), n15046900 (solid). ",
|
| 1553 |
+
"bbox": [
|
| 1554 |
+
174,
|
| 1555 |
+
867,
|
| 1556 |
+
823,
|
| 1557 |
+
924
|
| 1558 |
+
],
|
| 1559 |
+
"page_idx": 11
|
| 1560 |
+
},
|
| 1561 |
+
{
|
| 1562 |
+
"type": "table",
|
| 1563 |
+
"img_path": "images/2b85f116756f94af7fb634b44dfd9820263714f2b2e0678fa4b02bcae4b4283d.jpg",
|
| 1564 |
+
"table_caption": [
|
| 1565 |
+
"Table 5: Few-shot learning baseline results using labeled/unlabeled splits. Baselines either takes inputs directly from the pixel space or use a CNN to extract features. “rnd” denotes using a randomly initialized CNN, and “pre” denotes using a CNN that is pretrained for supervised classification for all training classes. "
|
| 1566 |
+
],
|
| 1567 |
+
"table_footnote": [],
|
| 1568 |
+
"table_body": "<table><tr><td rowspan=\"2\">Models</td><td>Omniglot</td><td colspan=\"2\">miniImageNet</td><td colspan=\"2\">tieredImageNet</td></tr><tr><td>1-shot</td><td>1-shot</td><td>5-shot</td><td>1-shot</td><td>5-shot</td></tr><tr><td>1-NN Pixel</td><td>40.39± 0.36</td><td>26.74 ± 0.48</td><td>31.43 ± 0.51</td><td>26.55± 0.50</td><td>30.79± 0.53</td></tr><tr><td>1-NN CNN rnd</td><td>59.55 ± 0.46</td><td>24.03 ± 0.38</td><td>27.54 ± 0.42</td><td>25.49 ± 0.45</td><td>30.01 ± 0.47</td></tr><tr><td>1-NN CNN pre</td><td>52.53 ± 0.51</td><td>32.90 ± 0.58</td><td>40.79 ± 0.76</td><td>32.76 ± 0.66</td><td>40.26 ± 0.67</td></tr><tr><td>LRPixel</td><td>49.15 ± 0.39</td><td>24.50 ± 0.41</td><td>33.33±0.68</td><td>25.70± 0.46</td><td>36.30± 0.62</td></tr><tr><td>LR CNN rnd</td><td>57.80 ± 0.45</td><td>24.10 ± 0.50</td><td>28.40 ± 0.42</td><td>26.55 ± 0.48</td><td>32.51 ± 0.52</td></tr><tr><td>LR CNN pre</td><td>48.49 ± 0.47</td><td>30.28 ± 0.54</td><td>40.27 ± 0.59</td><td>34.52 ± 0.68</td><td>43.58 ± 0.72</td></tr><tr><td>ProtoNet</td><td>94.62 ± 0.09</td><td>43.61 ± 0.27</td><td>59.08 ± 0.22</td><td>46.52 ± 0.32</td><td>66.15 ± 0.34</td></tr></table>",
|
| 1569 |
+
"bbox": [
|
| 1570 |
+
181,
|
| 1571 |
+
99,
|
| 1572 |
+
816,
|
| 1573 |
+
228
|
| 1574 |
+
],
|
| 1575 |
+
"page_idx": 12
|
| 1576 |
+
},
|
| 1577 |
+
{
|
| 1578 |
+
"type": "text",
|
| 1579 |
+
"text": "C EXTRA EXPERIMENTAL RESULTS ",
|
| 1580 |
+
"text_level": 1,
|
| 1581 |
+
"bbox": [
|
| 1582 |
+
176,
|
| 1583 |
+
301,
|
| 1584 |
+
486,
|
| 1585 |
+
318
|
| 1586 |
+
],
|
| 1587 |
+
"page_idx": 12
|
| 1588 |
+
},
|
| 1589 |
+
{
|
| 1590 |
+
"type": "text",
|
| 1591 |
+
"text": "C.1 FEW-SHOT CLASSIFICATION BASELINES ",
|
| 1592 |
+
"text_level": 1,
|
| 1593 |
+
"bbox": [
|
| 1594 |
+
178,
|
| 1595 |
+
333,
|
| 1596 |
+
490,
|
| 1597 |
+
347
|
| 1598 |
+
],
|
| 1599 |
+
"page_idx": 12
|
| 1600 |
+
},
|
| 1601 |
+
{
|
| 1602 |
+
"type": "text",
|
| 1603 |
+
"text": "We provide baseline results on few-shot classification using 1-nearest neighbor and logistic regression with either pixel inputs or CNN features. Compared with the baselines, Regular ProtoNet performs significantly better on all three few-shot classification datasets. ",
|
| 1604 |
+
"bbox": [
|
| 1605 |
+
174,
|
| 1606 |
+
358,
|
| 1607 |
+
825,
|
| 1608 |
+
400
|
| 1609 |
+
],
|
| 1610 |
+
"page_idx": 12
|
| 1611 |
+
},
|
| 1612 |
+
{
|
| 1613 |
+
"type": "text",
|
| 1614 |
+
"text": "C.2 NUMBER OF UNLABELED ITEMS ",
|
| 1615 |
+
"text_level": 1,
|
| 1616 |
+
"bbox": [
|
| 1617 |
+
176,
|
| 1618 |
+
417,
|
| 1619 |
+
439,
|
| 1620 |
+
431
|
| 1621 |
+
],
|
| 1622 |
+
"page_idx": 12
|
| 1623 |
+
},
|
| 1624 |
+
{
|
| 1625 |
+
"type": "text",
|
| 1626 |
+
"text": "Figure 6 shows test accuracy values with different number of unlabeled items during test time. \nFigure 7 shows our mask output value distribution of the Masked Soft $k$ -Means model on Omniglot. \nThe mask values have a bi-modal distribution, corresponding to distractor and non-distractor items. ",
|
| 1627 |
+
"bbox": [
|
| 1628 |
+
174,
|
| 1629 |
+
444,
|
| 1630 |
+
823,
|
| 1631 |
+
486
|
| 1632 |
+
],
|
| 1633 |
+
"page_idx": 12
|
| 1634 |
+
},
|
| 1635 |
+
{
|
| 1636 |
+
"type": "text",
|
| 1637 |
+
"text": "D HYPERPARAMETER DETAILS ",
|
| 1638 |
+
"text_level": 1,
|
| 1639 |
+
"bbox": [
|
| 1640 |
+
176,
|
| 1641 |
+
506,
|
| 1642 |
+
447,
|
| 1643 |
+
522
|
| 1644 |
+
],
|
| 1645 |
+
"page_idx": 12
|
| 1646 |
+
},
|
| 1647 |
+
{
|
| 1648 |
+
"type": "text",
|
| 1649 |
+
"text": "For Omniglot, we adopted the best hyperparameter settings found for ordinary Prototypical Networks in Snell et al. (2017). In these settings, the learning rate was set to 1e-3, and cut in half every 2K updates starting at update 2K. We trained for a total of 20K updates. For miniImagenet and tieredImageNet, we trained with a starting learning rate of 1e-3, which we also decayed. We started the decay after 25K updates, and every 25K updates thereafter we cut it in half. We trained for a total of 200K updates. We used ADAM (Kingma & Ba, 2014) for the optimization of our models. For the MLP used in the Masked Soft $k$ -Means model, we use a single hidden layer with 20 hidden units with a tanh non-linearity for all 3 datasets. We did not tune the hyparameters of this MLP so better performance may be attained with a more rigorous hyperparameter search. ",
|
| 1650 |
+
"bbox": [
|
| 1651 |
+
174,
|
| 1652 |
+
537,
|
| 1653 |
+
825,
|
| 1654 |
+
664
|
| 1655 |
+
],
|
| 1656 |
+
"page_idx": 12
|
| 1657 |
+
},
|
| 1658 |
+
{
|
| 1659 |
+
"type": "image",
|
| 1660 |
+
"img_path": "images/ffb830cb65e716257aac92f1eea002089be7b6cbf27a6e110c0024a441fb44be.jpg",
|
| 1661 |
+
"image_caption": [],
|
| 1662 |
+
"image_footnote": [],
|
| 1663 |
+
"bbox": [
|
| 1664 |
+
214,
|
| 1665 |
+
117,
|
| 1666 |
+
702,
|
| 1667 |
+
910
|
| 1668 |
+
],
|
| 1669 |
+
"page_idx": 13
|
| 1670 |
+
},
|
| 1671 |
+
{
|
| 1672 |
+
"type": "image",
|
| 1673 |
+
"img_path": "images/ee3b37c708ea4994f31dc10d74e3a6730fa8e3abdad49f0c128dfa6ea0a94972.jpg",
|
| 1674 |
+
"image_caption": [
|
| 1675 |
+
"Figure 6: Model Performance on tieredImageNet with different number of unlabeled items during test time. We include test accuracy numbers in this chart. "
|
| 1676 |
+
],
|
| 1677 |
+
"image_footnote": [],
|
| 1678 |
+
"bbox": [
|
| 1679 |
+
174,
|
| 1680 |
+
175,
|
| 1681 |
+
820,
|
| 1682 |
+
421
|
| 1683 |
+
],
|
| 1684 |
+
"page_idx": 14
|
| 1685 |
+
},
|
| 1686 |
+
{
|
| 1687 |
+
"type": "image",
|
| 1688 |
+
"img_path": "images/4ec598503b0edfd3103a0860e66835f81c8129fecdd7d502de22ec3016fed891.jpg",
|
| 1689 |
+
"image_caption": [
|
| 1690 |
+
"Figure 7: Mask values predicted by masked soft k-means on Omniglot. "
|
| 1691 |
+
],
|
| 1692 |
+
"image_footnote": [],
|
| 1693 |
+
"bbox": [
|
| 1694 |
+
285,
|
| 1695 |
+
622,
|
| 1696 |
+
714,
|
| 1697 |
+
813
|
| 1698 |
+
],
|
| 1699 |
+
"page_idx": 14
|
| 1700 |
+
}
|
| 1701 |
+
]
|
parse/train/HJcSzz-CZ/HJcSzz-CZ_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/HJcSzz-CZ/HJcSzz-CZ_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/S1xSSTNKDB/S1xSSTNKDB_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/S1xSSTNKDB/S1xSSTNKDB_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/Sy3fJXbA-/Sy3fJXbA-.md
ADDED
|
@@ -0,0 +1,282 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# CONNECTIVITY LEARNING IN MULTI-BRANCH NETWORKS
|
| 2 |
+
|
| 3 |
+
Anonymous authors Paper under double-blind review
|
| 4 |
+
|
| 5 |
+
# ABSTRACT
|
| 6 |
+
|
| 7 |
+
While much of the work in the design of convolutional networks over the last five years has revolved around the empirical investigation of the importance of depth, filter sizes, and number of feature channels, recent studies have shown that branching, i.e., splitting the computation along parallel but distinct threads and then aggregating their outputs, represents a new promising dimension for significant improvements in performance. To combat the complexity of design choices in multi-branch architectures, prior work has adopted simple strategies, such as a fixed branching factor, the same input being fed to all parallel branches, and an additive combination of the outputs produced by all branches at aggregation points. In this work we remove these predefined choices and propose an algorithm to learn the connections between branches in the network. Instead of being chosen a priori by the human designer, the multi-branch connectivity is learned simultaneously with the weights of the network by optimizing a single loss function defined with respect to the end task. We demonstrate our approach on the problem of multi-class image classification using four different datasets where it yields consistently higher accuracy compared to the state-of-the-art “ResNeXt” multi-branch network given the same learning capacity.
|
| 8 |
+
|
| 9 |
+
# 1 INTRODUCTION
|
| 10 |
+
|
| 11 |
+
Deep neural networks have emerged as one of the most prominent models for problems that require the learning of complex functions and that involve large amounts of training data. While deep learning has recently enabled dramatic performance improvements in many application domains, the design of deep architectures is still a challenging and time-consuming endeavor. The difficulty lies in the many architecture choices that impact—often significantly—the performance of the system. In the specific domain of image categorization, which is the focus of this paper, significant research effort has been invested in the empirical study of how depth, filter sizes, number of feature maps, and choice of nonlinearities affect performance (Glorot et al., 2011; Krizhevsky et al., 2012; Sermanet et al., 2013; Maas et al., 2013; Zeiler & Fergus, 2014; Szegedy et al., 2015). Recently, several authors have proposed to simplify the architecture design by defining convolutional neural networks (CNNs) in terms of combinations of basic building blocks. This strategy was arguably first popularized by the VGG networks (Simonyan & Zisserman, 2015) which were built by stacking a series of convolutional layers having identical filter size $\left( 3 \times 3 \right)$ . The idea of modularized CNN design was made even more explicit in residual networks (ResNets) (He et al., 2016), which are constructed by combining residual blocks of fixed topology. While in ResNets residual blocks are stacked one on top of each other to form very deep networks, the recently introduced ResNeXt models (Xie et al., 2017) have shown that it is also beneficial to arrange these building blocks in parallel to build multi-branch convolutional networks. The modular component of ResNeXt then consists of $C$ parallel branches, corresponding to residual blocks with identical topology but distinct parameters. Network built by stacking these multi-branch components have been shown to lead to better results than single-thread ResNets of the same capacity.
|
| 12 |
+
|
| 13 |
+
While the principle of modularized design has greatly simplified the challenge of building effective architectures for image analysis, the choice of how to combine and aggregate the computations of these building blocks still rests on the shoulders of the human designer. In order to avoid a combinatorial explosion of options, prior work has relied on simple, uniform rules of aggregation and composition. For example, ResNeXt models (Xie et al., 2017) are based on the following set of simplifying assumptions: the branching factor $C$ (also referred to as cardinality) is fixed to the mitted to 31st Conference on Neural Information Processing Systems (NIPS 2017). Do not distribute.same constant in all layers of the network, all branches of a module are fed the same input, and the outputs of parallel branches are aggregated by a simple additive operation that provides the input to the next module. In this paper we remove these predefined choices and propose an algorithm that Submitted to 31st Conference on Neural Information Processing Systems (NIPS 2017). Do not distlearns to combine and aggregate building blocks of a neural network. In this new regime, the network Submitted to 31st Conference on Neural Information Prconnectivity naturally arises as a result of the training optimization rather than being hand-defined by the human designer.
|
| 14 |
+
|
| 15 |
+

|
| 16 |
+
y = x + F (x; ✓ j ) (1)21 F (x; ✓2) 1Figure 1: Different types of building blocks for modular network design: (a) a prototypical residual j=1 1block with bottleneck convolutional layers (He et al., 2016); (b) the multi-branch RexNeXt module consisting of $C$ Cparallel residual blocks (Xie et al., 2017); (c) our approach replaces the fixed y = x + F(aggregation points of RexNeXt with learnable masks $\mathbf { m }$ ✓j ) (1)defining the input connections for each individual residual block.
|
| 17 |
+
|
| 18 |
+
itted to 31st Conference on Neural Information Processing Systems (NIPS 2017). Do noWe demonstrate our approach using residual blocks as our modular components, but we take inspiration from ResNeXt by arranging these modules in a multi-branch architecture. Rather than Submitted to 31st Conference on Neural Information Processing Systems (NIPS 2017). Do not distribute.predefining the input connections and aggregation pathways of each branch, we let the algorithm discover the optimal way to combine and connect residual blocks with respect to the end learning objective. This is achieved by means of masks, i.e., learned binary parameters that act as “switches” determining the final connectivity in our network. The masks are learned together with the convolutional weights of the network, as part of a joint optimization via backpropagation with respect to a traditional multi-class classification objective. We demonstrate that, given the same budget of residual blocks (and parameters), our learned architecture consistently outperforms the predefined ResNeXt network in all our experiments. An interesting byproduct of our approach is that it can automatically identify residual blocks that are superfluous, i.e., unnecessary or detrimental for the end objective. At the end of the optimization, these unused residual blocks can be pruned away without any impact on the learned hypothesis while yielding substantial savings in number of parameters to store and in test-time computation.
|
| 19 |
+
|
| 20 |
+
# 2 TECHNICAL APPROACH
|
| 21 |
+
|
| 22 |
+
# 2.1 MODULAR MULTI-BRANCH ARCHITECTURE
|
| 23 |
+
|
| 24 |
+
We begin by providing a brief review of residual blocks (He et al., 2016), which represent the modular components of our architecture. We then discuss ResNeXt (Xie et al., 2017), which inspired the multi-branch structure of our networks. Finally, we present our approach to learning the connectivity of multi-branch architectures using binary masks.
|
| 25 |
+
|
| 26 |
+
Residual Learning. The framework of residual learning was introduced by He et al. (He et al., 2016) as a strategy to cope with the challenging optimization of deep models. The approach was inspired by the observation that deeper neural networks, despite having larger learning capacity than shallower models, often yield higher training error, due to the difficulty of optimization posed by increasing depth. Yet, given any arbitrary shallow network, it is trivially possible to reproduce its function using a deeper model, e.g., by copying the shallow network into the top portion of the deep model and by setting the remaining layers to implement identity functions. This simple yet revealing intuition inspired the authors to introduce residual blocks, which learn residual functions with reference to the layer input. Figure 1(a) illustrates an example of these modular components where the 3 layers in the block implement a residual function $\mathcal F ( \mathbf x )$ . A shortcut connections aggregates the residual block output $\mathcal { F } ( \mathbf { x } )$ with its input $\mathbf { x }$ , thus computing $\mathcal { F } ( \mathbf { x } ) + \mathbf { x } .$ , which becomes the input to the next block. The point of this module is that if at any depth in the network the representation $\mathbf { x }$ is already optimal, then $\mathcal { F } ( \mathbf { x } )$ can be trivially set to be the zero function, which is easier to learn than an identity mapping. In fact, it was shown (He et al., 2016) that reformulating the layers as learning residuals eases optimization and enables the effective training of networks that are substantially deeper than previously possible. Since we are interested in applying our approach to image categorization, in this paper we use convolutional residual blocks using the bottleneck (He et al., 2016) shown in Figure 1(a). The first $1 \times 1$ layer projects the input feature maps onto a lower dimensional embedding, the second applies $3 \times 3$ filters, and the third restores the original feature map dimensionality. As in (He et al., 2016), Batch Normalization (Ioffe & Szegedy, 2015) and ReLU (Krizhevsky et al., 2012) are applied after each layer, and a ReLU is used after each aggregation.
|
| 27 |
+
|
| 28 |
+
The multi-branch architecture of ResNeXt. Recent work (Xie et al., 2017) has shown that it is beneficial to arrange residual blocks not only along the depth dimension but also to implement parallel multiple threads of computation feeding from the same input layer. The outputs of the parallel residual blocks are then summed up together with the original input and passed on to the next module. The resulting multi-branch module is illustrated in Figure 1(b). More formally, let $\mathcal { F } ( \mathbf { x } ; \theta _ { j } ^ { ( i ) } )$ be the transformation implemented by the $j$ -th residual block in module $i$ -th of the network, where $j = 1 , \ldots , C$ and $i = 1 , \ldots , L$ , with $L$ denoting the total number of modules stacked on top of each other to form the complete network. The hyperparameter $C$ is called the cardinality of the module and defines the number of parallel branches within each module. The hyperparameter $L$ controls the total depth of the network: under the assumption of 3 layers per residual block (as shown in the figure), the total depth of the network is given by $D = 2 + 3 L$ (an initial convolutional layer and an output fully-connected layers add 2 layers). Note that in ResNeXt all residual blocks in a module have the same topology $( \mathcal { F } )$ but each block has its own parameters $\cdot \theta _ { j } ^ { ( i ) }$ denotes the parameters of residual block $j$ in module $i$ ). Then, the output of the $i$ -th module is computed as:
|
| 29 |
+
|
| 30 |
+
$$
|
| 31 |
+
\mathbf { y } = \mathbf { x } + \sum _ { j = 1 } ^ { C } \mathcal { F } ( \mathbf { x } ; \boldsymbol { \theta } _ { j } ^ { ( i ) } )
|
| 32 |
+
$$
|
| 33 |
+
|
| 34 |
+
Tensor $\mathbf { y }$ represents the input to the $( i + 1 )$ -th module. Note that the ResNeXt module effectively implements a split-transform-merge strategy that perfoms a projection of the input into separate lower-dimensional embeddings (via bottlenecks), a separate transformation within each embedding, a projection back to the high-dimensional space and a final aggregation via addition. It can be shown that the solutions that can be implemented by such module are a strict subspace of the solutions of a single layer operating on the high-dimensional embedding but at a considerably lower cost in terms of computational complexity and number of parameters. In (Xie et al., 2017) it was experimentally shown that increasing the cardinality $C$ is a more effective way of improving accuracy compared to increasing depth or the number of filters. In other words, given a fixed budget of parameters, ResNeXt multi-branch networks were shown to consistently outperform single-branch ResNets of the same learning capacity.
|
| 35 |
+
|
| 36 |
+
We note, however, that in an attempt to ease network design, several restrictive limitations were embedded in the architecture of ResNeXt modules: each ResNeXt module implements $C$ parallel feature extractors that operate on the same input; furthermore, the number of active branches is constant at all depth levels of the network. In the next subsection we present an approach that removes these restrictions without adding any significant burden on the process of manual network design (with the exception of a single additional integer hyperparameter for the entire network).
|
| 37 |
+
|
| 38 |
+
Our masked multi-branch architecture. As in ResNeXt, our proposed architecture consists of a stack of $L$ multi-branch modules, each containing $C$ parallel feature extractors. However, differently from ResNeXt, each branch in a module can take a different input. The input pathway of each
|
| 39 |
+
|
| 40 |
+
branch is controlled by a binary mask vector that is learned jointly with the weights of the network.
|
| 41 |
+
Let con m(ij $\mathbf { m } _ { j } ^ { ( i ) } = [ m _ { j , 1 } ^ { ( i ) } , m _ { j , 2 } ^ { ( i ) } , \ldots , m _ { j , C } ^ { ( i ) } ] ^ { \top } \in \{ 0 , 1 \} ^ { C }$ be the binin module ry . If tor defining the active input, then the activation volume $j$ $i$ $m _ { j , k } ^ { ( i ) } = 1$
|
| 42 |
+
produced by the $k$ -th branch in module $( i - 1 )$ is fed as input to the $j$ -th residual block of module . If
|
| 43 |
+
$\mathbf { \dot { \rho } } _ { m _ { j , k } ^ { ( i ) } } = 0$ , then the output from the $k$ -th branch in the previous module is ignored by the $j$ -th residual
|
| 44 |
+
j,blothe of the current modul-th branch in module , if we den, the input with to t $\mathbf { \hat { y } } _ { k } ^ { ( i - 1 ) }$ the output activation tens residual block in module computed bywill be given $k$ $( i - 1 )$ $\mathbf { x } _ { j } ^ { ( i ) }$ $j$ $i$
|
| 45 |
+
by the following equation:
|
| 46 |
+
|
| 47 |
+
$$
|
| 48 |
+
\mathbf { x } _ { j } ^ { ( i ) } = \sum _ { k = 1 } ^ { C } m _ { j , k } ^ { ( i ) } \cdot \mathbf { y } _ { k } ^ { ( i - 1 ) }
|
| 49 |
+
$$
|
| 50 |
+
|
| 51 |
+
Then, the output of this block will be obtained through the usual residual computation, i.e., $\mathbf { y } _ { j } ^ { ( i ) } = \mathbf { x } _ { j } ^ { ( i ) } + \bar { \mathcal { F } } ( \mathbf { x } _ { j } ^ { ( i ) } ; \boldsymbol { \theta } _ { j } ^ { ( i ) } )$ . We note that under this model we no longer have fixed aggregation nodes summing up all outputs computed from a module. Instead, the mask $\mathbf { m } _ { j } ^ { ( i ) }$ now determines selectively for each block which branches from the previous module will be aggregated and provided as input to the block. Under this scheme, the parallel branches in a module receive different inputs and as such are likely to yield more diverse features.
|
| 52 |
+
|
| 53 |
+
We point out that depending on the constraints posed over $\mathbf { m } _ { i } ^ { ( i ) }$ , different interesting models can be realized. For example, by introducing the constraint that Pk m(i)j,k for all blocks $j$ , then each residual block will receive input from only one branch (since each m(i)j,k must be either 0 or 1). It can be noted that at the other end of the spectrum, if we set $\begin{array} { r } { \dot { m } _ { j , k } ^ { ( i ) } = 1 } \end{array}$ for all blocks $j , k$ in each module $i$ , then all connections would be active and we would obtain again the fixed ResNeXt architecture. In our experiments we will demonstrate that the best results are achieved for a middle ground between these two extremes, i.e., by connecting each block to $K$ branches where $K$ is an integer-valued hyperparameter such that $1 < K < C$ . We refer to this hyperparameter as the fan-in of a block. As discussed in the next section, the mask vector $\mathbf { m } _ { j } ^ { \left( i \right) }$ for each block is learned simultaneously with all the other weights in the network via backpropagation. Finally, we note that it may be possible for a residual block in the network to become unused. This happens when, as a result of the optimization, block k in module (i − 1) is such that m(i)jk $m _ { j k } ^ { ( i ) } = 0$ for all $j = 1 , \ldots , C$ . In this case, at the end of the optimization, we prune the block in order to reduce the number of parameters to store and to speed up inference (note that this does not affect the function computed by the network). Thus, at any point in the network the total number of active parallel threads can be any number smaller than or equal to $C$ . This implies that a variable branching factor is learned adaptively for the different depths in the network.
|
| 54 |
+
|
| 55 |
+
# 2.2 MASKCONNECT: LEARNING TO CONNECT BRANCHES
|
| 56 |
+
|
| 57 |
+
We refer to our learning algorithm as MASKCONNECT. It performs joint optimization of a given learning objective $\ell$ with respect to both the weights of the network $\mathbf { \eta } ^ { ( \theta ) }$ as well as the masks $\mathbf { \Pi } ^ { ( \mathbf { m } ) }$ . Since in this paper we apply our method to the problem of image categorization, we use the traditional multi-class cross-entropy objective for the loss $\ell$ . However, our approach can be applied without change to other loss functions as well as to other tasks benefitting from a multi-branch architecture.
|
| 58 |
+
|
| 59 |
+
In MASKCONNECT the weights have real values, as in traditional networks, while the branch masks have binary values. This renders the optimization more challenging. To learn these binary parameters, we adopt a modified version of backpropagation, inspired by the algorithm proposed by Courbariaux et al. (Courbariaux et al., 2015) to train neural networks with binary weights. During training we store and update a real-valued version $\tilde { \mathbf { m } } _ { j } ^ { ( i ) } \in [ 0 , 1 ] ^ { C }$ of the branch masks, with entries clipped to lie in the continuous interval from 0 to 1.
|
| 60 |
+
|
| 61 |
+
In general, the training via backpropagation consists of three steps: 1) forward propagation, 2) backward propagation, and 3) parameters update. At each iteration, we stochastically binarize the real-valued branch masks into binary-valued vectors $\mathbf { m } _ { j } ^ { ( i ) } \in \{ 0 , 1 \} ^ { \dot { C } }$ which are then used for the forward propagation and backward propagation (steps 1 and 2). Instead, during the parameters update (step 3), the method updates the real-valued branch masks $\tilde { \mathbf { m } } _ { j } ^ { ( i ) }$ . The weights $\theta$ of the convolutional and fully connected layers are optimized using standard backpropagation. We discuss below the details of our mask training procedure, under the constraint that at any time there can be only $K$ active entries in the binary branch mask $\mathbf { m } _ { j } ^ { ( i ) }$ , where $K$ is a predefined integer hyperparameter with
|
| 62 |
+
|
| 63 |
+
$1 \leq K \leq C$ . In other words, we impose the following constraints:
|
| 64 |
+
|
| 65 |
+
$$
|
| 66 |
+
m _ { j , k } ^ { ( i ) } \in \{ 0 , 1 \} , \quad \sum _ { k = 1 } ^ { C } m _ { j , k } ^ { ( i ) } = K \forall j \in \{ 1 , \ldots , C \} \mathrm { ~ a n d } \forall i \in \{ 1 , \ldots , L \} .
|
| 67 |
+
$$
|
| 68 |
+
|
| 69 |
+
These constraints imply that each residual block receives input from exactly $K$ branches of the previous module.
|
| 70 |
+
|
| 71 |
+
Forward Propagation. During the forreal-valued branch masks for each block (i) (i) (i) $j$ ard propagation, our algorithmto sum up to 1, i.e., such that $\begin{array} { r } { \sum _ { k = 1 } ^ { C } \tilde { m } _ { j , k } ^ { ( i ) } = 1 } \end{array}$ es the . This $C$ done so that Mult( ˜m j,1, m˜ j,2, . . , m˜ j,C ) defines a proper multinomial distribution over the $C$ branch connections feeding into block $j$ . Then, the binary branch mask $\mathbf { m } _ { j } ^ { ( i ) }$ is stochastically generated by drawing $K$ distinct samples $a _ { 1 } , a _ { 2 } , \ldots , a _ { K } \in \{ 1 , \ldots , C \}$ from the multinomial distribution over the branch connections. Finally, the entries corresponding to the $K$ samples are activated in the binary branch mask vector, i.e., m(i)j,ak $\dot { m } _ { j , a _ { k } } ^ { ( i ) } \gets 1$ , for $k = 1 , . . . , K$ . The input activation volume to the residual block sampl $j$ is then computed according to Eq. 2 from the sampled binary branch masks.g from the Multinomial distribution ensures that the connections with largest ote that thevalues will $\tilde { m } _ { j , k } ^ { ( i ) }$ be more likely to be chosen, while at the same time the stochasticity of this process allows different connectivities to be explored, particularly during early stages of the learning when the real-valued masks have still fairly uniform values.
|
| 72 |
+
|
| 73 |
+
Backward Propagation. In the backward propagation steto each branch output is obtained via back-propagation from adient and t $\partial \ell / \partial y _ { k } ^ { ( i - 1 ) }$ withasks ect. $\partial \ell / \partial x _ { j } ^ { \overline { { ( i ) } } }$ $m _ { j , k } ^ { ( i ) }$
|
| 74 |
+
|
| 75 |
+
Mask Update. In the parameter update step our algorithm computes the gradient with respect to the binary branch masks for each branch. Then, using these computed gradients and the given learning rate, it updates the real-valued branch masks via gradient descent. At this time we clip the updated real-valued branch masks to constrain them to remain within the valid interval $[ 0 , 1 ]$ . The same clipping strategy was adopted for the binary weights in the work of Courbariaux et al. (2015).
|
| 76 |
+
|
| 77 |
+
As discussed in the supplementary material, after joint training over $\theta$ and $\mathbf { m }$ , we have found beneficial to fine-tune the weights $\theta$ of the network with fixed binary masks (connectivity), by setting as active connections for each block $j$ in module $i$ those corresponding to the $K$ largest values in $\tilde { \mathbf { m } } _ { j } ^ { ( i ) }$ . Pseudocode for our training procedure is given in the supplementary material.
|
| 78 |
+
|
| 79 |
+
# 3 EXPERIMENTS
|
| 80 |
+
|
| 81 |
+
We tested our approach on the task of image categorization using several benchmarks: CIFAR10 (Krizhesvsky, 2009), CIFAR-100 (Krizhesvsky, 2009), Mini-ImageNet (Vinyals et al., 2016), as well as the full ImageNet (Deng et al., 2009). In this section we discuss results achieved on CIFAR-100 and ImageNet (Deng et al., 2009), while the results for CIFAR-10 (Krizhesvsky, 2009) and Mini-ImageNet (Vinyals et al., 2016) can be found in the Appendix.
|
| 82 |
+
|
| 83 |
+
# 3.1 CIFAR-100
|
| 84 |
+
|
| 85 |
+
CIFAR-100 is a dataset of color images of size $3 2 \mathbf { x } 3 2$ . It consists of 50,000 training images and 10,000 test images. Each image in CIFAR-100 is categorized into one of 100 possible classes.
|
| 86 |
+
|
| 87 |
+
Effect of fan-in $( K )$ . We start by studying the effect of the fan-in hyperparameter $( K )$ on the performance of models built and trained using our proposed approach. The fan-in defines the number of active branches feeding each residual block. For this experiment we use a model obtained by stacking $L = 6$ multi-branch residual modules, each having cardinality $C = 8$ (number of branches in each module). We use residual blocks consisting of 3 convolutional layers with a bottleneck implementing dimensionality reduction on the number of feature channels, as shown in Figure 1. The bottleneck for this experiment was set to $w = 4$ . Since each residual block consists of 3 layers, the total depth of the network in terms of learnable layers is $D = 2 + 3 L = 2 0$ .
|
| 88 |
+
|
| 89 |
+
We trained and tested this architecture using different fan-in values: $K = 1 , . . , 8$ . Note that varying $K$ does not affect the number of parameters. Thus, all these models have the same learning capacity.
|
| 90 |
+
|
| 91 |
+

|
| 92 |
+
Figure 2: Varying the fan-in $( K )$ of our model, i.e., the number of active branches provided as input to each residual block. The plot reports accuracy achieved on CIFAR-100 using a network stack of $L = 6$ ResNeXt modules having cardinality $C \ = \ 8$ and bottleneck width $w = 4$ . All models have the same number of parameters (0.28M). The best accuracy is obtained for $K = 4$ .
|
| 93 |
+
|
| 94 |
+

|
| 95 |
+
Figure 3: A visualization of the fixed branch connectivity of ResNext (left) versus the connectivity learned by our method (right) using $K = 1$ ). Each green square is a residual block, each row of $C \ = \ 8$ square is a multibranch module. The network consists of a stack of $L = 9$ modules. Arrows indicate pathways connecting residual blocks of adjacent modules. In each net, the top red circle is a convolutional layer, the bottom circle is the final fully-connected layer. It can be noticed that MASKCONNECT learns sparse connections. The squares without $_ \mathrm { i n / o u t }$ edges are those deemed superfluous by our algorithm and can be pruned at the end of learning. This gives rise to a branching factor that varies along the depth of the net.
|
| 96 |
+
|
| 97 |
+
The results are shown in Figure 2. We can see that the best accuracy is achieved by connecting each residual block to $K = 4$ branches out of the total $C = 8$ in each module. Using a very low or very high fan-in yields lower accuracy. Note that when setting $K = C$ , there is no need to learn the masks. In this case each mask is simply replaced by an element-wise addition of the outputs from all the branches. This renders the model equivalent to ResNeXt (Xie et al., 2017), which has fixed connectivity. Based on the results of Figure 2, in all our experiments below we use $K = 4$ , since it gives the best accuracy, but also $K = 1$ , since it gives high sparsity which, as we will see shortly, implies savings in number of parameters.
|
| 98 |
+
|
| 99 |
+
Varying the architectures. In Table 1 we show the classification accuracy achieved with different architectures (the details of each architecture are listed in the Appendix). For each architecture we report results obtained using MASKCONNECT with fan-in $K = 1$ and $K = 4$ . We also include the accuracy achieved with full (as opposed to learned) connectivity, which corresponds to ResNeXt. These results show that learning the connectivity produces consistently higher accuracy than using fixed connectivity, with accuracy gains of up $2 . 2 \%$ compared to the state-of-the-art ResNeXt model.
|
| 100 |
+
|
| 101 |
+
We note that these improvements in accuracy come at little computational training cost: the average training time overhead for learning masks and weights is about $3 9 \%$ using our unoptimized implementation compared to learning only the weights given a fixed connectivity. Additionally, for each architecture we include models trained using sparse random connectivity (Fixed-Random). For these models, each mask is set to have $K = 4$ randomly-chosen active connections, and the connectivity is kept fixed during learning of the parameters. We can notice that the accuracy of these nets is considerably lower compared to our models, despite having the same connectivity density $X = 4$ ). This shows that the improvements of our approach over ResNeXt are not due to sparser connectivity but they are rather due to learned connectivity.
|
| 102 |
+
|
| 103 |
+
Parameter savings. Our proposed approach provides the benefit of automatically identifying during training residual blocks that are unnecessary. At the end of the training, the unused residual blocks can be pruned away. This yields savings in the number of parameters to store and in testtime computation. In Table 1, columns Train and Test under Params show the original number of parameters (used during training) and the number of parameters after pruning (used at test-time). Note that for the biggest architecture, our approach using $K = 1$ yields a parameter saving of $40 \%$ compared to ResNeXt with full connectivity (20.5M vs 34.4M), while achieving the same accuracy. Thus, in summary, using fan-in $K = 4$ gives models that have the same number of parameters as ResNeXt but they yield higher accuracy; using fan-in $K = 1$ gives a significant saving in number of parameters and accuracy on par with ResNeXt.
|
| 104 |
+
|
| 105 |
+
Model with real-valued masks. We have also attempted to learn our models using real-valued masks by computing tensors in the forward and backward propagation with respect to masks $\tilde { \mathbf { m } } _ { j } ^ { ( i ) } \in [ \tilde { 0 , 1 } ] ^ { C }$ rather than the binary vectors $\mathbf { m } _ { j } ^ { ( i ) } \in \{ 0 , 1 \} ^ { C }$ . However, we found this variant to
|
| 106 |
+
|
| 107 |
+
Table 1: CIFAR-100 accuracies (single crop) achieved by different architectures trained using the predefined full connectivity of ResNeXt (Fixed-Full) versus the connectivity learned by our algorithm (Learned). We also include models trained using random, fixed connectivity (Fixed-Random) defined by setting $K = 4$ random active connections per branch. Each model was trained 4 times, using different random initializations. For each model we report the best test performance as well as the mean test performance computed from the 4 runs. For our method, we report performance using $K = 1$ as well as $K = 4$ . We also list the number of parameters used during training (Params-Train) and the number of parameters obtained after pruning the unused blocks (Params-Test). Our learned connectivity using $K = 4$ produces accuracy gains of up $2 . 2 \%$ compared to the strong ResNeXt model, while using $K = 1$ yields results equivalent to ResNeXt but it induces a significant reduction in number of parameters at test time (a saving of $40 \%$ for model $\{ 2 9 , 6 4 , 8 \} ,$ ).
|
| 108 |
+
|
| 109 |
+
<table><tr><td rowspan="2">Architecture {Depth (D), Bottleneck width (w),Cardinality (C)}</td><td rowspan="2">Connectivity</td><td colspan="2">Params</td><td>Accuracy (%)</td></tr><tr><td>Train</td><td>Test</td><td>Top-1 best (mean±std)</td></tr><tr><td rowspan="4">{29,8,8}</td><td>Fixed-Full, K=8 (Xie et al.,2017)</td><td>0.86M</td><td>0.86M</td><td>73.52 (73.37±0.13)</td></tr><tr><td>Learned, K=1</td><td>0.86M</td><td>0.65M</td><td>73.91 (73.76±0.14)</td></tr><tr><td>Learned, K=4</td><td>0.86M</td><td>0.81M</td><td>75.89 (75.77±0.12)</td></tr><tr><td>Fixed-Random, K=4</td><td>0.86M</td><td>0.85M</td><td>72.85 (72.66±0.24)</td></tr><tr><td rowspan="4">{29,64,8}</td><td>Fixed-Full, K=8 (Xie et al.,2017)</td><td>34.4M</td><td>34.4M</td><td>82.23 (82.12±0.12)</td></tr><tr><td>Learned, K=1</td><td>34.4M</td><td>20.5M</td><td>82.31 (82.15±0.15)</td></tr><tr><td>Learned, K=4</td><td>34.4M</td><td>32.1M</td><td>84.05 (83.94±0.11)</td></tr><tr><td>Fixed-Random, K=4</td><td>34.4M</td><td>34.3M</td><td>81.96 (81.73±0.20)</td></tr></table>
|
| 110 |
+
|
| 111 |
+
yield consistently lower results compared to our models using binary masks. For example, for model $\{ D = 2 9 , \stackrel { \cdot } { w } = 8 , C = 8 \}$ the best accuracy achieved with real-valued masks is $1 . 9 3 \%$ worse compared to that obtained with binary masks. In particular we observed that for this variant, the real-valued masks change little over training even when using large learning rates. Conversely, performing the forward and backward propagation using stochastically-sampled binary masks yields a larger exploration of connectivities and results in bigger changes of the auxiliary real-valued masks leading to better connectivity learning.
|
| 112 |
+
|
| 113 |
+
Visualization of the learned connectivity. Figure 3 provides an illustration of the connectivity learned by MASKCONNECT for $K = 1$ versus the fixed connectivity of ResNeXt for model $\{ D =$ 2 $9 , w = 8 , C = 8 \}$ . While ResNeXt feeds the same input to all blocks of a module, our algorithm learns different input pathways for each block and yields a branching factor that varies along depth.
|
| 114 |
+
|
| 115 |
+
# 3.2 IMAGENET
|
| 116 |
+
|
| 117 |
+
Finally, we evaluate our approach on the large-scale ImageNet 2012 dataset (Deng et al., 2009), which includes images of 1000 classes. We train our approach on the training set (1.28M images) and evaluate it on the validation set (50K images). In Table 2, we report the Top-1 and Top-5 accuracies for three different architectures. For these experiments we set $\bar { K } = C / 2$ . We can observe that for all three architectures, our learned connectivity yields an improvement in accuracy over the fixed connectivity of ResNeXt (Xie et al., 2017).
|
| 118 |
+
|
| 119 |
+
# 3.3 CIFAR-10 & MINI-IMAGENET
|
| 120 |
+
|
| 121 |
+
We invite the reader to review results achieved on CIFAR-10 & Mini-ImageNet in the Appendix. Also on these datasets our algorithm consistently outperforms the ResNeXt models based on fixed connectivity, with accuracy gains of up to $3 . 8 \%$ .
|
| 122 |
+
|
| 123 |
+
# 4 RELATED WORK
|
| 124 |
+
|
| 125 |
+
Despite their wide adoption, deep networks often require laborious model search in order to yield good results. As a result, significant research effort has been devoted to the design of algorithms for automatic model selection. However, most of this prior work falls within the genre of hyperparameter optimization (Bergstra & Bengio, 2012; Snoek et al., 2012; 2015) rather than architecture or connectivity learning. Evolutionary search has been proposed as an interesting framework to learn both the structure as well as the connections in a neural network (Wierstra et al., 2005; Floreano et al., 2008; Real et al., 2017). Architecture search has also been recently formulated as a reinforcement learning problem with impressive results (Zoph & Le, 2017). Unlike these approaches, our method is limited to learning the connectivity within a predefined architecture but it does so efficiently by gradient descent optimization of the learning objective as opposed to more costly procedures such as evolutionary search or reinforcement learning. Several authors have proposed learning connectivity by pruning unimportant weights from the network (LeCun et al., 1989; Han et al., 2015a;b; Guo et al., 2016; Han et al., 2016). However, these prior methods operate in stages where initially the network with full connectivity is learned and then connections are greedily removed according to an importance criterion. In PathNet (Fernando et al., 2017), the connectivity within a given architecture was searched via evolution. Compare to these prior approaches, our work provides the advantage of learning the connectivity by direct global optimization of the loss function of the problem at hand rather than by greedy optimization of a proxy criterion or by evolution. Our technical approach shares similarities with the “Shake-Shake” regularization recently introduced in unpublished work (Gastaldi, 2017). This procedure was demonstrated on two-branch ResNeXt models and consists in randomly scaling tensors produced by parallel branches during each training iteration while at test time the network uses uniform weighting of tensors. Conversely, our algorithm learns an optimal binary scaling of the parallel tensors with respect to the training objective and uses the resulting network with sparse connectivity at test time. Our work is also related to approaches that learn a hierarchical structure in the last one or two layers of a network in order to obtain distinct features for different categories (Murdock et al., 2016; Ahmed & Torresani, 2017). Differently from these methods, our algorithm learns efficiently connections at all depths in the network, thus optimizing over a much larger family of connectivity models. While our algorithm is limited to optimizing the connectivity structure within a predefined architecture, Adams et al. (Adams et al., 2010) proposed a nonparametric Bayesian approach that searches over an infinite network using MCMC. Saxena and Verbeek (Saxena & Verbeek, 2016) introduced convolutional neural fabric which are learnable 3D trellises that locally connect response maps at different layers of a CNN. Similarly to our work, they enable optimization over an exponentially large family of connectivities, albeit different from those considered here.
|
| 126 |
+
|
| 127 |
+
Table 2: ImageNet accuracies (single crop) achieved by different architectures using the predefined connectivity of ResNeXt (Fixed-Full) versus the connectivity learned by our algorithm (Learned).
|
| 128 |
+
|
| 129 |
+
<table><tr><td>Architecture</td><td>Connectivity</td><td colspan="2">Accuracy</td></tr><tr><td>{Depth (D), Bottleneck width (w), Cardinality (C}</td><td></td><td>Top-1</td><td>Top-5</td></tr><tr><td rowspan="2">{50,4,32}</td><td>Fixed-Full, K=32 (Xie et al., 2017)</td><td>77.8</td><td>93.3</td></tr><tr><td>Learned, K=16</td><td>79.1</td><td>94.1</td></tr><tr><td rowspan="2">{101,4,32}</td><td>Fixed-Full, K=32 (Xie et al.,2017)</td><td>78.8</td><td>94.1</td></tr><tr><td>Learned, K=16</td><td>79.5</td><td>94.5</td></tr><tr><td rowspan="2">{101,4,64}</td><td>Fixed-Full, K=64 (Xie et al., 2017)</td><td>79.6</td><td>94.7</td></tr><tr><td>Learned, K=32</td><td>79.8</td><td>94.8</td></tr></table>
|
| 130 |
+
|
| 131 |
+
# 5 CONCLUSIONS
|
| 132 |
+
|
| 133 |
+
In this paper we introduced an algorithm to learn the connectivity of deep multi-branch networks. The problem is formulated as a single joint optimization over the weights and the branch connections of the model. We tested our approach on challenging image categorization benchmarks where it led to significant accuracy improvements over the state-of-the-art ResNeXt model. An added benefit of our approach is that it can automatically identify superfluous blocks, which can be pruned without impact on accuracy for more efficient testing and for reducing the number of parameters to store.
|
| 134 |
+
|
| 135 |
+
While our experiments were focused on a particular multi-branch architecture (ResNeXt) and a specific form of building block (residual block), we expect the benefits of our approach to extend to other modules and network structures. For example, it could be applied to learn the connectivity of skip-connections in DenseNets (Huang et al., 2017), which are currently based on predefined connectivity rules. In this paper, our masks perform non-parametric additive aggregation of the branch outputs. It would be interesting to experiment with learnable (parametric) aggregations of the outputs from the individual branches. Our approach is limited to learning connectivity within a given, fixed architecture. Future work will explore the use of learnable masks for architecture discovery.
|
| 136 |
+
|
| 137 |
+
# REFERENCES
|
| 138 |
+
|
| 139 |
+
Ryan Prescott Adams, Hanna M. Wallach, and Zoubin Ghahramani. Learning the structure of deep sparse graphical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, AISTATS 2010, Chia Laguna Resort, Sardinia, Italy, May 13-15, 2010, pp. 1–8, 2010.
|
| 140 |
+
|
| 141 |
+
Karim Ahmed and Lorenzo Torresani. Branchconnect: Large-scale visual recognition with learned branch connections. CoRR, abs/1704.06010, 2017. URL http://arxiv.org/abs/1704. 06010.
|
| 142 |
+
|
| 143 |
+
James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13:281–305, 2012.
|
| 144 |
+
|
| 145 |
+
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in Neural Information Processing Systems 28, Montreal, Quebec, Canada, pp. 3123–3131, 2015.
|
| 146 |
+
|
| 147 |
+
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pp. 248–255, 2009.
|
| 148 |
+
|
| 149 |
+
Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. CoRR, abs/1701.08734, 2017. URL http://arxiv.org/abs/1701.08734.
|
| 150 |
+
|
| 151 |
+
Dario Floreano, Peter Durr, and Claudio Mattiussi. Neuroevolution: from architectures to learning. ¨ Evolutionary Intelligence, 1(1):47–62, 2008.
|
| 152 |
+
|
| 153 |
+
Xavier Gastaldi. Shake-shake regularization. CoRR, abs/1705.07485, 2017. URL http://arxiv. org/abs/1705.07485.
|
| 154 |
+
|
| 155 |
+
Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, AISTATS 2011, Fort Lauderdale, USA, April 11-13, 2011, pp. 315–323, 2011.
|
| 156 |
+
|
| 157 |
+
Yiwen Guo, Anbang Yao, and Yurong Chen. Dynamic network surgery for efficient dnns. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pp. 1379–1387, 2016.
|
| 158 |
+
|
| 159 |
+
Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding. In International Conference on Learning Representations (ICLR), 2015a.
|
| 160 |
+
|
| 161 |
+
Song Han, Jeff Pool, John Tran, and William J. Dally. Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems 28, Montreal, Quebec, Canada, pp. 1135–1143, 2015b.
|
| 162 |
+
|
| 163 |
+
Song Han, Jeff Pool, Sharan Narang, Huizi Mao, Shijian Tang, Erich Elsen, Bryan Catanzaro, John Tran, and William J. Dally. DSD: regularizing deep neural networks with dense-sparse-dense training flow. In International Conference on Learning Representations (ICLR), 2016.
|
| 164 |
+
|
| 165 |
+
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), 2016 IEEE Conference on, 2016.
|
| 166 |
+
|
| 167 |
+
Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, 2017.
|
| 168 |
+
|
| 169 |
+
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pp. 448–456, 2015.
|
| 170 |
+
|
| 171 |
+
Alex Krizhesvsky. Learning multiple layers of features from tiny images, 2009. Technical Report https://www.cs.toronto.edu/˜kriz/learning-features-2009-TR.pdf.
|
| 172 |
+
|
| 173 |
+
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25, Lake Tahoe, Nevada, United States., pp. 1106–1114, 2012.
|
| 174 |
+
|
| 175 |
+
Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-30, 1989], pp. 598–605, 1989.
|
| 176 |
+
|
| 177 |
+
Andrew L Maas, Awni Y Hannun, and Andrew Y Ng. Rectifier nonlinearities improve neural network acoustic models. Proc. ICML, 30:1, 2013.
|
| 178 |
+
|
| 179 |
+
Calvin Murdock, Zhen Li, Howard Zhou, and Tom Duerig. Blockout: Dynamic model selection for hierarchical deep networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pp. 2583–2591, 2016.
|
| 180 |
+
|
| 181 |
+
Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR), 2017.
|
| 182 |
+
|
| 183 |
+
Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Quoc V. Le, and Alex Kurakin. Large-scale evolution of image classifiers. CoRR, abs/1703.01041, 2017.
|
| 184 |
+
|
| 185 |
+
Shreyas Saxena and Jakob Verbeek. Convolutional neural fabrics. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pp. 4053–4061, 2016.
|
| 186 |
+
|
| 187 |
+
Pierre Sermanet, David Eigen, Xiang Zhang, Michael Mathieu, Rob Fergus, and Yann LeCun. ¨ Overfeat: Integrated recognition, localization and detection using convolutional networks. In International Conference on Learning Representations (ICLR), 2013.
|
| 188 |
+
|
| 189 |
+
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015.
|
| 190 |
+
|
| 191 |
+
Jasper Snoek, Hugo Larochelle, and Ryan P. Adams. Practical bayesian optimization of machine learning algorithms. In Advances in Neural Information Processing Systems 25, Lake Tahoe, Nevada, United States., pp. 2960–2968, 2012.
|
| 192 |
+
|
| 193 |
+
Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan Kiros, Nadathur Satish, Narayanan Sundaram, Md. Mostofa Ali Patwary, Prabhat, and Ryan P. Adams. Scalable bayesian optimization using deep neural networks. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pp. 2171–2180, 2015.
|
| 194 |
+
|
| 195 |
+
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pp. 1–9, 2015.
|
| 196 |
+
|
| 197 |
+
Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems 29, Barcelona, Spain, pp. 3630–3638, 2016.
|
| 198 |
+
|
| 199 |
+
Daan Wierstra, Faustino J. Gomez, and Jurgen Schmidhuber. Modeling systems with internal state ¨ using evolino. In Genetic and Evolutionary Computation Conference, GECCO 2005, Proceedings, Washington DC, USA, June 25-29, 2005, pp. 1795–1802, 2005.
|
| 200 |
+
|
| 201 |
+
Saining Xie, Ross B. Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual ´ transformations for deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, 2017.
|
| 202 |
+
|
| 203 |
+
Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I, pp. 818–833, 2014.
|
| 204 |
+
|
| 205 |
+
Barret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning. In International Conference on Learning Representations (ICLR), 2017.
|
| 206 |
+
|
| 207 |
+
# A PSEUDOCODE OF THE ALGORITHM
|
| 208 |
+
|
| 209 |
+
# Algorithm 1 MASKCONNECT training algorithm.
|
| 210 |
+
|
| 211 |
+
Input: a minibatch of labeled examples $( x ^ { i } , y ^ { i } )$ , $C$ : cardinality (number of branches), $K$ : fan-in (number of active branch connections), $\eta$ : learning rate, $\ell$ : the loss over the minibatch, $\tilde { \mathbf { m } } _ { j } ^ { ( i ) } \in [ 0 , 1 ] ^ { C }$ : real-valued branch masks for block $j$ in module $_ { i }$ from previous training iteration.
|
| 212 |
+
Output: updated $\tilde { \mathbf { m } } _ { j } ^ { ( i ) }$
|
| 213 |
+
1. Forward Propagation:
|
| 214 |
+
Normalize the real-valued mask to sum up to 1: m˜ (i)j,k $\begin{array} { r } { \tilde { m } _ { j , k } ^ { ( i ) } \gets \frac { \tilde { m } _ { j , k } ^ { ( i ) } } { \sum _ { k ^ { \prime } = 1 } ^ { C } \tilde { m } _ { j , k ^ { \prime } } ^ { ( i ) } } } \end{array}$ , for $j = 1 , \ldots , C$
|
| 215 |
+
Reset binary mask: $\mathbf { m } _ { j } ^ { ( i ) } \gets \mathbf { 0 }$
|
| 216 |
+
Draw $K$ distinct samples from multinomial mask distribution:
|
| 217 |
+
$a _ { 1 } , a _ { 2 } , \ldots , a _ { K } \gets \hat { \mathbf { M u l t } } ( \tilde { m } _ { j , 1 } ^ { ( i ) } , \tilde { m } _ { j , 2 } ^ { ( i ) } , \ldots , \tilde { m } _ { j , C } ^ { ( i ) } )$
|
| 218 |
+
Set active binary mask based on drawn samples:
|
| 219 |
+
$m _ { j , a _ { k } } ^ { ( i ) } \gets 1$ for $k = 1 , . . . , K$
|
| 220 |
+
Compute output $\mathbf { x } _ { j } ^ { ( i ) }$ of the mask, given branch activations $\begin{array} { r } { \mathbf { y } _ { k } ^ { ( i - 1 ) } \colon \mathbf { x } _ { j } ^ { ( i ) } \sum _ { k = 1 } ^ { C } m _ { j , k } ^ { ( i ) } \cdot \mathbf { y } _ { k } ^ { ( i - 1 ) } } \end{array}$
|
| 221 |
+
j2. Backward Propagation:
|
| 222 |
+
Compute Compute $\frac { \partial \ell } { \partial \mathbf { x } _ { j } ^ { ( i ) } }$ om fro ∂ y(i)j \`
|
| 223 |
+
$\frac { \partial \ell } { \partial \mathbf { y } _ { k } ^ { ( i - 1 ) } }$ ∂x(i)j and m(i)j,k
|
| 224 |
+
3. Parameter Update:
|
| 225 |
+
Compute m˜ (i) ← c $\frac { \partial \ell } { \partial m _ { j , k } ^ { ( i ) } }$ given (i) − η ∂ x(i)j d y(ik an\` ) $\mathbf { y } _ { k } ^ { ( i - 1 ) }$
|
| 226 |
+
$\begin{array} { r } { \tilde { m } _ { j , k } ^ { ( i ) } \mathrm { c l i p } ( \tilde { m } _ { j , k } ^ { ( i ) } - \eta \cdot \frac { \partial \ell } { \partial m _ { j , k } ^ { ( i ) } } ) } \end{array}$
|
| 227 |
+
|
| 228 |
+
# B EXPERIMENTS ON CIFAR-10
|
| 229 |
+
|
| 230 |
+
The CIFAR-10 dataset consists of color images of size $3 2 \mathrm { x } 3 2$ . The training set contains 50,000 images, the testing set 10,000 images. Each image in CIFAR-10 is categorized into one of 10 possible classes. In Table 3, we report the performance of different models trained on CIFAR-10. From these results we can observe that our models using learned connectivity achieve consistently better performance over the equivalent models trained with the fixed connectivity (Xie et al., 2017).
|
| 231 |
+
|
| 232 |
+
Table 3: CIFAR-10 accuracies (single crop) achieved by different multi-branch architectures trained using the predefined connectivity of ResNeXt (Fixed-Full) versus the connectivity learned by our algorithm (Learned). Each model was trained 4 times, using different random initializations. For each model we report the best test performance as well as the mean test performance computed from the 4 runs.
|
| 233 |
+
|
| 234 |
+
<table><tr><td>Architecture</td><td>Connectivity</td><td>Accuracy (%)</td></tr><tr><td>{Depth (D), Bottleneck width (w), Cardinality (C)}</td><td></td><td>Top-1 best (mean±std)</td></tr><tr><td>{20,4,8}</td><td>Fixed-Full K=8 (Xie et al., 2017) Learned K=4</td><td>91.39 (91.13±0.11) 92.85 (92.76±0.10)</td></tr><tr><td>{29,4,8}</td><td>Fixed-Full K=8 (Xie et al.,2017) Learned K=4</td><td>92.77 (92.65±0.09) 93.88 (93.76±0.12)</td></tr><tr><td>{29,8.8}</td><td>Fixed-Full K=8 (Xie et al.,2017) Learned K=4</td><td>93.26 (93.14±0.11) 95.11 (94.96±0.12)</td></tr><tr><td>{29,64,8}</td><td>Fixed-Full K=8 (Xie et al.,2017) Learned K=4</td><td>96.35 (96.23±0.12) 96.83 (96.73±0.11)</td></tr></table>
|
| 235 |
+
|
| 236 |
+
Table 4: Mini-ImageNet accuracies achieved by different multi-branch networks trained using the predefined full connectivity of ResNeXt (Fixed-Full) versus the connectivity learned by our algorithm (Learned). Additionally, we include models trained using random fixed connectivity (Fixed-Random) for $K = 4$ . For each model we report the best and the mean test performance computed from 4 different training runs. Our method for joint learning of weights and connectivity yields a gain of over $3 \%$ in Top-1 accuracy over ResNeXt, which uses the same architectures but a fixed branch connectivity.
|
| 237 |
+
|
| 238 |
+
<table><tr><td>Architecture</td><td>Connectivity</td><td>Accuracy</td></tr><tr><td>{Depth (D), Bottleneck width (w), Cardinality (C}</td><td></td><td>Top-1 best (mean±std)</td></tr><tr><td rowspan="3">{20,4,8}</td><td>Fixed-Full K=8 (Xie et al., 2017)</td><td>62.12 (61.86±0.15)</td></tr><tr><td>Learned K=4</td><td>66.09 (65.94±0.16)</td></tr><tr><td>Fixed-Random K=4</td><td>62.42 (61.81±0.32)</td></tr><tr><td rowspan="3">{29,8,8}</td><td>Fixed-Full K=8 (Xie et al., 2017)</td><td>68.11 (67.89±0.19)</td></tr><tr><td>Learned K=4</td><td>71.36 (71.18±0.19)</td></tr><tr><td>Fixed-Random K=4</td><td>67.97 (67.53±0.20)</td></tr></table>
|
| 239 |
+
|
| 240 |
+
# C EXPERIMENTS ON MINI-IMAGENET
|
| 241 |
+
|
| 242 |
+
Mini-ImageNet is a subset of the full ImageNet (Deng et al., 2009) dataset. It was used in (Vinyals et al., 2016; Ravi & Larochelle, 2017). It is created by randomly selecting 100 classes from the full ImageNet (Deng et al., 2009). For each class, 600 images are randomly selected. We use 500 examples per class for training, and the other 100 examples per class for testing. The selected images are resized to size 84x84 pixels as in (Vinyals et al., 2016; Ravi & Larochelle, 2017). The advantage of this dataset is that it poses the recognition challenges typical of the ImageNet photos but at the same time it does not need require the powerful resources needed to train on the full ImageNet dataset. This allows to include the additional baselines involving random fixed connectivity (Fixed-Random).
|
| 243 |
+
|
| 244 |
+
We report the performance of different models trained on Mini-ImageNet in Table 4. From these results, we see that our models using learned connectivity with fan-in $K { = } 4$ yield a nice accuracy gain over the same models trained with the fixed full connectivity of ResNeXt (Xie et al., 2017). The absolute improvement (in Top-1 accuracy) is $3 . 8 7 \%$ for the 20-layer network and $3 . 1 7 \%$ for the 29-layer network. We can notice that the accuracy of the models with fixed random connectivity (Fixed-Random) is considerably lower compared to our nets with learned connectivity, despite having the same connectivity density $K = 4$ ). This shows that the improvement of our approach over ResNeXt is not due to sparser connectivity but it is rather due to learned connectivity.
|
| 245 |
+
|
| 246 |
+
# D VISUALIZATIONS OF LEARNED CONNECTIVITY
|
| 247 |
+
|
| 248 |
+
The plot in Figure 4 shows how the number of active branches varies as a function of the module depth for model $\{ D = 2 9 , w = 4 , C = 8 \}$ trained on CIFAR-100. For $K = 1$ , we can observe that the number of active branches tends to be larger for deep modules (closer to the output layer) compared to early modules (closer to the input). We observed this phenomenon consistently for all architectures. This suggests that having many parallel threads of computation is particularly important in deep layers of the network. Conversely, the setting $K = 4$ tends to produce a fairly uniform number of active branches across the modules and the number is quite close to the maximum value $C$ . For this reason, there is little saving in terms of number of parameters when using $K = 4$ , as there are rarely unused blocks.
|
| 249 |
+
|
| 250 |
+
The plot in Figure 5 shows the number of active branches as a function of module depth for model $\{ D \stackrel { - } { = } 5 0 , w \stackrel { - } { = } 4 , C = 3 2 \}$ trained on ImageNet, using $K = 1 6$ .
|
| 251 |
+
|
| 252 |
+

|
| 253 |
+
Figure 4: Number of active branches as a function of module depth for model $\{ D = 2 9 , w = 4 , C =$ $8 \}$ trained on CIFAR-100. We report how the number of active branches varies for model trained with fan-in $K = 1$ as well as for the net trained with $K = 4$ . The setting $K = 1$ tends to leave many blocks unused, especially in the early modules of the network.
|
| 254 |
+
|
| 255 |
+

|
| 256 |
+
Figure 5: Number of active branches as a function of module depth for model $\{ D = 5 0 , w = 4 , C =$ $\mathrm { 3 2 } \bar \}$ trained on ImageNet, using fan-in $K = 1 6$ .
|
| 257 |
+
|
| 258 |
+
Table 5: Specifications of the architectures used in our experiments on the CIFAR-10 and CIFAR-100 datasets. The architectures differ in terms of depth $( D )$ , bottleneck width $( w )$ , and cardinality $( C )$ . Inside the brackets we specify the residual block used in each multi-branch module by listing the number of input channels, the size of the convolutional filters, as well as the number of filters (number of output channels). To the right of each bracket we list the cardinality (i.e., the number of parallel branches in the module). $\times 2$ means that the same multi-branch module is stacked twice. The first layer for all models is a convolutional layer with 16 filters of size $3 \times 3$ . The last layer performs global average pooling followed by a softmax.
|
| 259 |
+
|
| 260 |
+
<table><tr><td colspan="3">{D=20,w=4,C=8}</td><td colspan="3">{D=29,w=4,C=8}</td><td colspan="3">{D=29,w=8,C=8}</td><td colspan="2">{D=29,w=64,C=8}</td></tr><tr><td colspan="3">3,3×3,16</td><td colspan="2">3,3×3,16</td><td colspan="3">3,3×3,16</td><td colspan="3">3,3×3,64</td></tr><tr><td>16,1×1,4 4,3×3,4 [4,1×1,64 [64,1×1,4] 4,3×3,4</td><td>(C=8) (C=8)</td><td></td><td>[16,1×1,4 4,3×3,4 [4,1×1,64 [64,1×1,4]</td><td>(C=8)</td><td>16,1×1,8 8,3×3,8 8,1×1,64 [64,1×1,8]</td><td>(C=8)</td><td></td><td>64,1×1,64 64,3×3,64 64,1×1,256 [256,1×1,64</td><td>(C=8)</td></tr><tr><td>[4,1×1,64 [64,1×1,8 8,3×3,8 [8,1×1,128 [128,1×1,8 8,3×3,8</td><td></td><td>(C=8)</td><td>4,3×3,4 [4,1×1,64 [64,1×1,8 8,3×3,8 [8,1×1,128</td><td>(C=8),×2 (C=8)</td><td>8,3×3,8 [8,1×1,64] [64,1×1,16 16,3×3,16 16,1×1,128</td><td>(C=8),×2 (C=8)</td><td>64,3×3,64</td><td>64,1×1,256 256,1×1,128 128,3×3,128 128,1×1,512</td><td>(C=8),×2 (C=8)</td></tr><tr><td>8,1×1,128 128,1×1,16</td><td></td><td>(C=8)</td><td>[128,1×1,8 8,3×3,8 [8,1×1,128 128,1×1,16</td><td>(C=8),×2</td><td>[128,1×1,16 16,3×3,16 16,1×1,128 128,1×1,32</td><td>(C=8),×2</td><td></td><td>[512,1×1,128] 128,3×3,128 128,1×1,512] 512,1×1,256</td><td>(C=8), ×2</td></tr><tr><td colspan="2">16,3×3,16 [16,1×1,256 [256,1×1,16] 16,3×3,16 16,1×1,256</td><td>(C=8) (C=8)</td><td colspan="2">16,3×3,16 (C=8) [16,1×1,256 [256,1×1,16] 16,3×3,16 (C=8),×2</td><td>32,3×3,32 32,1×1,256 256,1×1,32 32,3×3,32</td><td>(C=8)</td><td></td><td>256,3×3,256 [256,1×1,1024 [1024,1×1,256 256,3×3,256</td><td>(C=8)</td></tr><tr><td colspan="2">Average Pool 100 fc, softmax</td><td></td><td colspan="2">[16,1×1,256 Average Pool 100 fc,softmax</td><td>32,1×1,256 100 fc, softmax</td><td>Average Pool</td><td>(C=8),×2</td><td colspan="2">(C=8),×2 256,1×1,1024 Average Pool</td></tr></table>
|
| 261 |
+
|
| 262 |
+
# E IMPLEMENTATION DETAILS
|
| 263 |
+
|
| 264 |
+
E.1 ARCHITECTURES AND SETTINGS FOR EXPERIMENTS ON CIFAR-100 AND CIFAR-10
|
| 265 |
+
|
| 266 |
+
The specifications of the architectures used in all our experiments on CIFAR-10 and CIFAR-100 are given in Table 5.
|
| 267 |
+
|
| 268 |
+
Several of these architectures are those presented in the original ResNeXt paper (Xie et al., 2017) and are trained using the same setup, including the data augmentation strategy.Four pixels are padded on each side of the input image, and a $3 2 \mathbf { x } 3 2$ crop is randomly sampled from the padded image or its horizontal flip, with per-pixel mean subtracted (Krizhevsky et al., 2012). For testing, we use the original $3 2 \mathbf { x } 3 2$ image. The stacks have output feature map of size 32, 16, and 8 respectively. The models are trained on 8 GPUs with a mini-batch size of 128 (16 per GPU), with a weight decay of 0.0005 and momentum of 0.9. We adopt four incremental training phases with a total of 320 epochs. In phase 1 we train the model for 120 epochs with a learning rate of 0.1 for the convolutional and fully-connected layers, and a learning rate of 0.2 for the masks. In phase 2 we freeze the connectivity by setting as active connections for each block those corresponding to its top- $K$ values in the masks. With these fixed learned connectivity, we finetune the model from phase $^ { l }$ for 100 epochs with a learning rate of 0.1 for the weights. Then, in phase 3 we finetune the weights of the model from phase 2 for 50 epochs with a learning rate of 0.01 using again the fixed learned connectivity from phase 1. Finally, in phase 4 we finetune the weights of the model from phase 3 for 50 epochs with a learning rate of 0.001.
|
| 269 |
+
|
| 270 |
+
# E.2 ARCHITECTURES AND SETTINGS FOR EXPERIMENTS ON IMAGENET
|
| 271 |
+
|
| 272 |
+
The architectures for our ImageNet experiments are those specified in the original ResNeXt paper (Xie et al., 2017).
|
| 273 |
+
|
| 274 |
+
Table 6: Mini-ImageNet architectures with varying depth $( D )$ , and bottleneck width $( w )$ . Inside the brackets we specify the residual block used in each multi-branch module by listing the number of input channels, the size of the convolutional filters, as well as the number of filters (number of output channels). To the right of each bracket we list the cardinality $( C )$ (i.e., the number of parallel branches in the module). $\times 2$ means that the same multi-branch module is stacked twice.
|
| 275 |
+
|
| 276 |
+
<table><tr><td colspan="2">{D=20,w=4,C=8}</td><td colspan="2">{D=29,=8,C=8}</td></tr><tr><td colspan="2">3,3×3,16</td><td colspan="2">3,3×3,16</td></tr><tr><td colspan="2">Max Pool,3×3,stride=2</td><td colspan="2">Max Pool,3×3,stride=2</td></tr><tr><td colspan="2">[16,1×1,4 4,3×3,4 (C=8) [4,1×1,64 [64,1×1,4 4,3×3,4 (C=8) [4,1×1,64</td><td colspan="2">[16,1×1,8 8,3×3,8 (C=8) [8,1×1,64 [64,1×1,8] 8,3×3,8 (C=8),×2</td></tr><tr><td colspan="2">[64,1×1,8 8,3×3,8 [8,1×1,128 [128,1×1,8 8,3×3,8</td><td colspan="2">[64,1×1,16 (C=8) 16,3×3,16 16,1×1,128 [128,1×1,16 (C=8) 16,3×3,16</td></tr><tr><td colspan="2">[8,1×1,128 [128,1×1,16 16,3×3,16 (C=8)</td><td colspan="2">(C=8),×2 [16,1×1,128] 128,1×1,32</td></tr><tr><td colspan="2">16,1×1,256 [256,1×1,16] 16,3×3,16 (C=8) 16,1×1,256</td><td colspan="2">[32,1×1,256 [256,1×1,32] 32,3×3,32 (C=8),×2 32,1×1,256]</td></tr><tr><td colspan="2">Average Pool 100 fc, softmax</td><td colspan="2">Average Pool 100 fc,softmax</td></tr></table>
|
| 277 |
+
|
| 278 |
+
Also for these experiments, we follow the data augmentation strategy described in (Xie et al., 2017). The input image has size $2 2 4 \mathbf { x } 2 2 4$ and it is randomly cropped from the resized original image. We use a mini-batch size of 256 on 8 GPUs (32 per GPU), with a weight decay of 0.0001 and a momentum of 0.9. We use four incremental training phases with a total of 120 epochs. In phase 1 we train the model for 30 epochs with a learning rate of 0.1 for the convolutional and fully-connected layers, and a learning rate of 0.2 for the masks. In phase 2 we finetune the model from phase $^ { l }$ for another 30 epochs with a learning rate of 0.1 and a learning rate of 0.0 for the masks (i.e., we use the fixed connectivity learned in phase 1). In phase 3 we finetune the weights from phase 2 for 30 epochs with a learning rate of 0.01 and the learning rate of the masks is 0.0. Finally, in phase 4 we finetune the weights from phase 3 for 30 epochs with a learning rate of 0.001 while the learning rate of the masks is still set to 0.0.
|
| 279 |
+
|
| 280 |
+
# E.3 ARCHITECTURES AND SETTINGS FOR EXPERIMENTS ON MINI-IMAGENET
|
| 281 |
+
|
| 282 |
+
For the experiments on the Mini-ImageNet dataset, a 64x64 crop is randomly sampled from the scaled 84x84 image or its horizontal flip, with per-pixel mean subtracted (Krizhevsky et al., 2012). For testing, we use the center $6 4 \mathrm { x } 6 4$ crop. The specifications of the models are identical to the CIFAR-100 models used in the previous subsection, except that the first input convolutional layer in the network is followed by a max pooling layer. The models are trained on 8 GPUs with a mini-batch size of 256 (32 per GPU), with a weight decay of 0.0005 and momentum of 0.9. Similar to training CIFAR-100 dataset, we also adopt four incremental training phases with a total of 320 epochs.
|
parse/train/Sy3fJXbA-/Sy3fJXbA-_content_list.json
ADDED
|
@@ -0,0 +1,1470 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "CONNECTIVITY LEARNING IN MULTI-BRANCH NETWORKS ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
174,
|
| 8 |
+
99,
|
| 9 |
+
553,
|
| 10 |
+
146
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Anonymous authors Paper under double-blind review ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
184,
|
| 19 |
+
171,
|
| 20 |
+
398,
|
| 21 |
+
198
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "ABSTRACT ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
454,
|
| 31 |
+
234,
|
| 32 |
+
544,
|
| 33 |
+
251
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "While much of the work in the design of convolutional networks over the last five years has revolved around the empirical investigation of the importance of depth, filter sizes, and number of feature channels, recent studies have shown that branching, i.e., splitting the computation along parallel but distinct threads and then aggregating their outputs, represents a new promising dimension for significant improvements in performance. To combat the complexity of design choices in multi-branch architectures, prior work has adopted simple strategies, such as a fixed branching factor, the same input being fed to all parallel branches, and an additive combination of the outputs produced by all branches at aggregation points. In this work we remove these predefined choices and propose an algorithm to learn the connections between branches in the network. Instead of being chosen a priori by the human designer, the multi-branch connectivity is learned simultaneously with the weights of the network by optimizing a single loss function defined with respect to the end task. We demonstrate our approach on the problem of multi-class image classification using four different datasets where it yields consistently higher accuracy compared to the state-of-the-art “ResNeXt” multi-branch network given the same learning capacity. ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
232,
|
| 42 |
+
267,
|
| 43 |
+
766,
|
| 44 |
+
507
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "1 INTRODUCTION ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
176,
|
| 54 |
+
536,
|
| 55 |
+
334,
|
| 56 |
+
553
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "Deep neural networks have emerged as one of the most prominent models for problems that require the learning of complex functions and that involve large amounts of training data. While deep learning has recently enabled dramatic performance improvements in many application domains, the design of deep architectures is still a challenging and time-consuming endeavor. The difficulty lies in the many architecture choices that impact—often significantly—the performance of the system. In the specific domain of image categorization, which is the focus of this paper, significant research effort has been invested in the empirical study of how depth, filter sizes, number of feature maps, and choice of nonlinearities affect performance (Glorot et al., 2011; Krizhevsky et al., 2012; Sermanet et al., 2013; Maas et al., 2013; Zeiler & Fergus, 2014; Szegedy et al., 2015). Recently, several authors have proposed to simplify the architecture design by defining convolutional neural networks (CNNs) in terms of combinations of basic building blocks. This strategy was arguably first popularized by the VGG networks (Simonyan & Zisserman, 2015) which were built by stacking a series of convolutional layers having identical filter size $\\left( 3 \\times 3 \\right)$ . The idea of modularized CNN design was made even more explicit in residual networks (ResNets) (He et al., 2016), which are constructed by combining residual blocks of fixed topology. While in ResNets residual blocks are stacked one on top of each other to form very deep networks, the recently introduced ResNeXt models (Xie et al., 2017) have shown that it is also beneficial to arrange these building blocks in parallel to build multi-branch convolutional networks. The modular component of ResNeXt then consists of $C$ parallel branches, corresponding to residual blocks with identical topology but distinct parameters. Network built by stacking these multi-branch components have been shown to lead to better results than single-thread ResNets of the same capacity. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
569,
|
| 66 |
+
825,
|
| 67 |
+
861
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "While the principle of modularized design has greatly simplified the challenge of building effective architectures for image analysis, the choice of how to combine and aggregate the computations of these building blocks still rests on the shoulders of the human designer. In order to avoid a combinatorial explosion of options, prior work has relied on simple, uniform rules of aggregation and composition. For example, ResNeXt models (Xie et al., 2017) are based on the following set of simplifying assumptions: the branching factor $C$ (also referred to as cardinality) is fixed to the mitted to 31st Conference on Neural Information Processing Systems (NIPS 2017). Do not distribute.same constant in all layers of the network, all branches of a module are fed the same input, and the outputs of parallel branches are aggregated by a simple additive operation that provides the input to the next module. In this paper we remove these predefined choices and propose an algorithm that Submitted to 31st Conference on Neural Information Processing Systems (NIPS 2017). Do not distlearns to combine and aggregate building blocks of a neural network. In this new regime, the network Submitted to 31st Conference on Neural Information Prconnectivity naturally arises as a result of the training optimization rather than being hand-defined by the human designer. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
867,
|
| 77 |
+
823,
|
| 78 |
+
924
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "image",
|
| 84 |
+
"img_path": "images/a5214bbd9af958c12bdcad52a245c78635be65faf57dc10a332288fabeda7482.jpg",
|
| 85 |
+
"image_caption": [
|
| 86 |
+
"y = x + F (x; ✓ j ) (1)21 F (x; ✓2) 1Figure 1: Different types of building blocks for modular network design: (a) a prototypical residual j=1 1block with bottleneck convolutional layers (He et al., 2016); (b) the multi-branch RexNeXt module consisting of $C$ Cparallel residual blocks (Xie et al., 2017); (c) our approach replaces the fixed y = x + F(aggregation points of RexNeXt with learnable masks $\\mathbf { m }$ ✓j ) (1)defining the input connections for each individual residual block. "
|
| 87 |
+
],
|
| 88 |
+
"image_footnote": [],
|
| 89 |
+
"bbox": [
|
| 90 |
+
176,
|
| 91 |
+
102,
|
| 92 |
+
823,
|
| 93 |
+
310
|
| 94 |
+
],
|
| 95 |
+
"page_idx": 1
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"type": "text",
|
| 99 |
+
"text": "",
|
| 100 |
+
"bbox": [
|
| 101 |
+
174,
|
| 102 |
+
426,
|
| 103 |
+
825,
|
| 104 |
+
537
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "itted to 31st Conference on Neural Information Processing Systems (NIPS 2017). Do noWe demonstrate our approach using residual blocks as our modular components, but we take inspiration from ResNeXt by arranging these modules in a multi-branch architecture. Rather than Submitted to 31st Conference on Neural Information Processing Systems (NIPS 2017). Do not distribute.predefining the input connections and aggregation pathways of each branch, we let the algorithm discover the optimal way to combine and connect residual blocks with respect to the end learning objective. This is achieved by means of masks, i.e., learned binary parameters that act as “switches” determining the final connectivity in our network. The masks are learned together with the convolutional weights of the network, as part of a joint optimization via backpropagation with respect to a traditional multi-class classification objective. We demonstrate that, given the same budget of residual blocks (and parameters), our learned architecture consistently outperforms the predefined ResNeXt network in all our experiments. An interesting byproduct of our approach is that it can automatically identify residual blocks that are superfluous, i.e., unnecessary or detrimental for the end objective. At the end of the optimization, these unused residual blocks can be pruned away without any impact on the learned hypothesis while yielding substantial savings in number of parameters to store and in test-time computation. ",
|
| 111 |
+
"bbox": [
|
| 112 |
+
173,
|
| 113 |
+
545,
|
| 114 |
+
826,
|
| 115 |
+
739
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "2 TECHNICAL APPROACH ",
|
| 122 |
+
"text_level": 1,
|
| 123 |
+
"bbox": [
|
| 124 |
+
176,
|
| 125 |
+
762,
|
| 126 |
+
401,
|
| 127 |
+
777
|
| 128 |
+
],
|
| 129 |
+
"page_idx": 1
|
| 130 |
+
},
|
| 131 |
+
{
|
| 132 |
+
"type": "text",
|
| 133 |
+
"text": "2.1 MODULAR MULTI-BRANCH ARCHITECTURE ",
|
| 134 |
+
"text_level": 1,
|
| 135 |
+
"bbox": [
|
| 136 |
+
174,
|
| 137 |
+
795,
|
| 138 |
+
516,
|
| 139 |
+
809
|
| 140 |
+
],
|
| 141 |
+
"page_idx": 1
|
| 142 |
+
},
|
| 143 |
+
{
|
| 144 |
+
"type": "text",
|
| 145 |
+
"text": "We begin by providing a brief review of residual blocks (He et al., 2016), which represent the modular components of our architecture. We then discuss ResNeXt (Xie et al., 2017), which inspired the multi-branch structure of our networks. Finally, we present our approach to learning the connectivity of multi-branch architectures using binary masks. ",
|
| 146 |
+
"bbox": [
|
| 147 |
+
174,
|
| 148 |
+
821,
|
| 149 |
+
825,
|
| 150 |
+
877
|
| 151 |
+
],
|
| 152 |
+
"page_idx": 1
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"type": "text",
|
| 156 |
+
"text": "Residual Learning. The framework of residual learning was introduced by He et al. (He et al., 2016) as a strategy to cope with the challenging optimization of deep models. The approach was inspired by the observation that deeper neural networks, despite having larger learning capacity than shallower models, often yield higher training error, due to the difficulty of optimization posed by increasing depth. Yet, given any arbitrary shallow network, it is trivially possible to reproduce its function using a deeper model, e.g., by copying the shallow network into the top portion of the deep model and by setting the remaining layers to implement identity functions. This simple yet revealing intuition inspired the authors to introduce residual blocks, which learn residual functions with reference to the layer input. Figure 1(a) illustrates an example of these modular components where the 3 layers in the block implement a residual function $\\mathcal F ( \\mathbf x )$ . A shortcut connections aggregates the residual block output $\\mathcal { F } ( \\mathbf { x } )$ with its input $\\mathbf { x }$ , thus computing $\\mathcal { F } ( \\mathbf { x } ) + \\mathbf { x } .$ , which becomes the input to the next block. The point of this module is that if at any depth in the network the representation $\\mathbf { x }$ is already optimal, then $\\mathcal { F } ( \\mathbf { x } )$ can be trivially set to be the zero function, which is easier to learn than an identity mapping. In fact, it was shown (He et al., 2016) that reformulating the layers as learning residuals eases optimization and enables the effective training of networks that are substantially deeper than previously possible. Since we are interested in applying our approach to image categorization, in this paper we use convolutional residual blocks using the bottleneck (He et al., 2016) shown in Figure 1(a). The first $1 \\times 1$ layer projects the input feature maps onto a lower dimensional embedding, the second applies $3 \\times 3$ filters, and the third restores the original feature map dimensionality. As in (He et al., 2016), Batch Normalization (Ioffe & Szegedy, 2015) and ReLU (Krizhevsky et al., 2012) are applied after each layer, and a ReLU is used after each aggregation. ",
|
| 157 |
+
"bbox": [
|
| 158 |
+
174,
|
| 159 |
+
895,
|
| 160 |
+
823,
|
| 161 |
+
924
|
| 162 |
+
],
|
| 163 |
+
"page_idx": 1
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"type": "text",
|
| 167 |
+
"text": "",
|
| 168 |
+
"bbox": [
|
| 169 |
+
173,
|
| 170 |
+
103,
|
| 171 |
+
825,
|
| 172 |
+
367
|
| 173 |
+
],
|
| 174 |
+
"page_idx": 2
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"type": "text",
|
| 178 |
+
"text": "The multi-branch architecture of ResNeXt. Recent work (Xie et al., 2017) has shown that it is beneficial to arrange residual blocks not only along the depth dimension but also to implement parallel multiple threads of computation feeding from the same input layer. The outputs of the parallel residual blocks are then summed up together with the original input and passed on to the next module. The resulting multi-branch module is illustrated in Figure 1(b). More formally, let $\\mathcal { F } ( \\mathbf { x } ; \\theta _ { j } ^ { ( i ) } )$ be the transformation implemented by the $j$ -th residual block in module $i$ -th of the network, where $j = 1 , \\ldots , C$ and $i = 1 , \\ldots , L$ , with $L$ denoting the total number of modules stacked on top of each other to form the complete network. The hyperparameter $C$ is called the cardinality of the module and defines the number of parallel branches within each module. The hyperparameter $L$ controls the total depth of the network: under the assumption of 3 layers per residual block (as shown in the figure), the total depth of the network is given by $D = 2 + 3 L$ (an initial convolutional layer and an output fully-connected layers add 2 layers). Note that in ResNeXt all residual blocks in a module have the same topology $( \\mathcal { F } )$ but each block has its own parameters $\\cdot \\theta _ { j } ^ { ( i ) }$ denotes the parameters of residual block $j$ in module $i$ ). Then, the output of the $i$ -th module is computed as: ",
|
| 179 |
+
"bbox": [
|
| 180 |
+
173,
|
| 181 |
+
382,
|
| 182 |
+
825,
|
| 183 |
+
577
|
| 184 |
+
],
|
| 185 |
+
"page_idx": 2
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"type": "equation",
|
| 189 |
+
"img_path": "images/011a41405aa7722516513a070cef0768d7d6b21fda35fb456eda5648597e2c42.jpg",
|
| 190 |
+
"text": "$$\n\\mathbf { y } = \\mathbf { x } + \\sum _ { j = 1 } ^ { C } \\mathcal { F } ( \\mathbf { x } ; \\boldsymbol { \\theta } _ { j } ^ { ( i ) } )\n$$",
|
| 191 |
+
"text_format": "latex",
|
| 192 |
+
"bbox": [
|
| 193 |
+
418,
|
| 194 |
+
575,
|
| 195 |
+
578,
|
| 196 |
+
619
|
| 197 |
+
],
|
| 198 |
+
"page_idx": 2
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"type": "text",
|
| 202 |
+
"text": "Tensor $\\mathbf { y }$ represents the input to the $( i + 1 )$ -th module. Note that the ResNeXt module effectively implements a split-transform-merge strategy that perfoms a projection of the input into separate lower-dimensional embeddings (via bottlenecks), a separate transformation within each embedding, a projection back to the high-dimensional space and a final aggregation via addition. It can be shown that the solutions that can be implemented by such module are a strict subspace of the solutions of a single layer operating on the high-dimensional embedding but at a considerably lower cost in terms of computational complexity and number of parameters. In (Xie et al., 2017) it was experimentally shown that increasing the cardinality $C$ is a more effective way of improving accuracy compared to increasing depth or the number of filters. In other words, given a fixed budget of parameters, ResNeXt multi-branch networks were shown to consistently outperform single-branch ResNets of the same learning capacity. ",
|
| 203 |
+
"bbox": [
|
| 204 |
+
174,
|
| 205 |
+
622,
|
| 206 |
+
825,
|
| 207 |
+
776
|
| 208 |
+
],
|
| 209 |
+
"page_idx": 2
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"type": "text",
|
| 213 |
+
"text": "We note, however, that in an attempt to ease network design, several restrictive limitations were embedded in the architecture of ResNeXt modules: each ResNeXt module implements $C$ parallel feature extractors that operate on the same input; furthermore, the number of active branches is constant at all depth levels of the network. In the next subsection we present an approach that removes these restrictions without adding any significant burden on the process of manual network design (with the exception of a single additional integer hyperparameter for the entire network). ",
|
| 214 |
+
"bbox": [
|
| 215 |
+
174,
|
| 216 |
+
784,
|
| 217 |
+
823,
|
| 218 |
+
867
|
| 219 |
+
],
|
| 220 |
+
"page_idx": 2
|
| 221 |
+
},
|
| 222 |
+
{
|
| 223 |
+
"type": "text",
|
| 224 |
+
"text": "Our masked multi-branch architecture. As in ResNeXt, our proposed architecture consists of a stack of $L$ multi-branch modules, each containing $C$ parallel feature extractors. However, differently from ResNeXt, each branch in a module can take a different input. The input pathway of each ",
|
| 225 |
+
"bbox": [
|
| 226 |
+
176,
|
| 227 |
+
882,
|
| 228 |
+
823,
|
| 229 |
+
924
|
| 230 |
+
],
|
| 231 |
+
"page_idx": 2
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"type": "text",
|
| 235 |
+
"text": "branch is controlled by a binary mask vector that is learned jointly with the weights of the network. \nLet con m(ij $\\mathbf { m } _ { j } ^ { ( i ) } = [ m _ { j , 1 } ^ { ( i ) } , m _ { j , 2 } ^ { ( i ) } , \\ldots , m _ { j , C } ^ { ( i ) } ] ^ { \\top } \\in \\{ 0 , 1 \\} ^ { C }$ be the binin module ry . If tor defining the active input, then the activation volume $j$ $i$ $m _ { j , k } ^ { ( i ) } = 1$ \nproduced by the $k$ -th branch in module $( i - 1 )$ is fed as input to the $j$ -th residual block of module . If \n$\\mathbf { \\dot { \\rho } } _ { m _ { j , k } ^ { ( i ) } } = 0$ , then the output from the $k$ -th branch in the previous module is ignored by the $j$ -th residual \nj,blothe of the current modul-th branch in module , if we den, the input with to t $\\mathbf { \\hat { y } } _ { k } ^ { ( i - 1 ) }$ the output activation tens residual block in module computed bywill be given $k$ $( i - 1 )$ $\\mathbf { x } _ { j } ^ { ( i ) }$ $j$ $i$ \nby the following equation: ",
|
| 236 |
+
"bbox": [
|
| 237 |
+
173,
|
| 238 |
+
103,
|
| 239 |
+
825,
|
| 240 |
+
217
|
| 241 |
+
],
|
| 242 |
+
"page_idx": 3
|
| 243 |
+
},
|
| 244 |
+
{
|
| 245 |
+
"type": "equation",
|
| 246 |
+
"img_path": "images/e08a4bfa56510c940781795de770a000d232a12fc93e678e1431342980196b18.jpg",
|
| 247 |
+
"text": "$$\n\\mathbf { x } _ { j } ^ { ( i ) } = \\sum _ { k = 1 } ^ { C } m _ { j , k } ^ { ( i ) } \\cdot \\mathbf { y } _ { k } ^ { ( i - 1 ) }\n$$",
|
| 248 |
+
"text_format": "latex",
|
| 249 |
+
"bbox": [
|
| 250 |
+
416,
|
| 251 |
+
219,
|
| 252 |
+
580,
|
| 253 |
+
263
|
| 254 |
+
],
|
| 255 |
+
"page_idx": 3
|
| 256 |
+
},
|
| 257 |
+
{
|
| 258 |
+
"type": "text",
|
| 259 |
+
"text": "Then, the output of this block will be obtained through the usual residual computation, i.e., $\\mathbf { y } _ { j } ^ { ( i ) } = \\mathbf { x } _ { j } ^ { ( i ) } + \\bar { \\mathcal { F } } ( \\mathbf { x } _ { j } ^ { ( i ) } ; \\boldsymbol { \\theta } _ { j } ^ { ( i ) } )$ . We note that under this model we no longer have fixed aggregation nodes summing up all outputs computed from a module. Instead, the mask $\\mathbf { m } _ { j } ^ { ( i ) }$ now determines selectively for each block which branches from the previous module will be aggregated and provided as input to the block. Under this scheme, the parallel branches in a module receive different inputs and as such are likely to yield more diverse features. ",
|
| 260 |
+
"bbox": [
|
| 261 |
+
173,
|
| 262 |
+
270,
|
| 263 |
+
825,
|
| 264 |
+
353
|
| 265 |
+
],
|
| 266 |
+
"page_idx": 3
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"type": "text",
|
| 270 |
+
"text": "We point out that depending on the constraints posed over $\\mathbf { m } _ { i } ^ { ( i ) }$ , different interesting models can be realized. For example, by introducing the constraint that Pk m(i)j,k for all blocks $j$ , then each residual block will receive input from only one branch (since each m(i)j,k must be either 0 or 1). It can be noted that at the other end of the spectrum, if we set $\\begin{array} { r } { \\dot { m } _ { j , k } ^ { ( i ) } = 1 } \\end{array}$ for all blocks $j , k$ in each module $i$ , then all connections would be active and we would obtain again the fixed ResNeXt architecture. In our experiments we will demonstrate that the best results are achieved for a middle ground between these two extremes, i.e., by connecting each block to $K$ branches where $K$ is an integer-valued hyperparameter such that $1 < K < C$ . We refer to this hyperparameter as the fan-in of a block. As discussed in the next section, the mask vector $\\mathbf { m } _ { j } ^ { \\left( i \\right) }$ for each block is learned simultaneously with all the other weights in the network via backpropagation. Finally, we note that it may be possible for a residual block in the network to become unused. This happens when, as a result of the optimization, block k in module (i − 1) is such that m(i)jk $m _ { j k } ^ { ( i ) } = 0$ for all $j = 1 , \\ldots , C$ . In this case, at the end of the optimization, we prune the block in order to reduce the number of parameters to store and to speed up inference (note that this does not affect the function computed by the network). Thus, at any point in the network the total number of active parallel threads can be any number smaller than or equal to $C$ . This implies that a variable branching factor is learned adaptively for the different depths in the network. ",
|
| 271 |
+
"bbox": [
|
| 272 |
+
173,
|
| 273 |
+
359,
|
| 274 |
+
825,
|
| 275 |
+
597
|
| 276 |
+
],
|
| 277 |
+
"page_idx": 3
|
| 278 |
+
},
|
| 279 |
+
{
|
| 280 |
+
"type": "text",
|
| 281 |
+
"text": "2.2 MASKCONNECT: LEARNING TO CONNECT BRANCHES ",
|
| 282 |
+
"text_level": 1,
|
| 283 |
+
"bbox": [
|
| 284 |
+
174,
|
| 285 |
+
618,
|
| 286 |
+
575,
|
| 287 |
+
632
|
| 288 |
+
],
|
| 289 |
+
"page_idx": 3
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"type": "text",
|
| 293 |
+
"text": "We refer to our learning algorithm as MASKCONNECT. It performs joint optimization of a given learning objective $\\ell$ with respect to both the weights of the network $\\mathbf { \\eta } ^ { ( \\theta ) }$ as well as the masks $\\mathbf { \\Pi } ^ { ( \\mathbf { m } ) }$ . Since in this paper we apply our method to the problem of image categorization, we use the traditional multi-class cross-entropy objective for the loss $\\ell$ . However, our approach can be applied without change to other loss functions as well as to other tasks benefitting from a multi-branch architecture. ",
|
| 294 |
+
"bbox": [
|
| 295 |
+
174,
|
| 296 |
+
645,
|
| 297 |
+
825,
|
| 298 |
+
715
|
| 299 |
+
],
|
| 300 |
+
"page_idx": 3
|
| 301 |
+
},
|
| 302 |
+
{
|
| 303 |
+
"type": "text",
|
| 304 |
+
"text": "In MASKCONNECT the weights have real values, as in traditional networks, while the branch masks have binary values. This renders the optimization more challenging. To learn these binary parameters, we adopt a modified version of backpropagation, inspired by the algorithm proposed by Courbariaux et al. (Courbariaux et al., 2015) to train neural networks with binary weights. During training we store and update a real-valued version $\\tilde { \\mathbf { m } } _ { j } ^ { ( i ) } \\in [ 0 , 1 ] ^ { C }$ of the branch masks, with entries clipped to lie in the continuous interval from 0 to 1. ",
|
| 305 |
+
"bbox": [
|
| 306 |
+
174,
|
| 307 |
+
722,
|
| 308 |
+
825,
|
| 309 |
+
805
|
| 310 |
+
],
|
| 311 |
+
"page_idx": 3
|
| 312 |
+
},
|
| 313 |
+
{
|
| 314 |
+
"type": "text",
|
| 315 |
+
"text": "In general, the training via backpropagation consists of three steps: 1) forward propagation, 2) backward propagation, and 3) parameters update. At each iteration, we stochastically binarize the real-valued branch masks into binary-valued vectors $\\mathbf { m } _ { j } ^ { ( i ) } \\in \\{ 0 , 1 \\} ^ { \\dot { C } }$ which are then used for the forward propagation and backward propagation (steps 1 and 2). Instead, during the parameters update (step 3), the method updates the real-valued branch masks $\\tilde { \\mathbf { m } } _ { j } ^ { ( i ) }$ . The weights $\\theta$ of the convolutional and fully connected layers are optimized using standard backpropagation. We discuss below the details of our mask training procedure, under the constraint that at any time there can be only $K$ active entries in the binary branch mask $\\mathbf { m } _ { j } ^ { ( i ) }$ , where $K$ is a predefined integer hyperparameter with ",
|
| 316 |
+
"bbox": [
|
| 317 |
+
174,
|
| 318 |
+
811,
|
| 319 |
+
825,
|
| 320 |
+
924
|
| 321 |
+
],
|
| 322 |
+
"page_idx": 3
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"type": "text",
|
| 326 |
+
"text": "$1 \\leq K \\leq C$ . In other words, we impose the following constraints: ",
|
| 327 |
+
"bbox": [
|
| 328 |
+
173,
|
| 329 |
+
103,
|
| 330 |
+
606,
|
| 331 |
+
118
|
| 332 |
+
],
|
| 333 |
+
"page_idx": 4
|
| 334 |
+
},
|
| 335 |
+
{
|
| 336 |
+
"type": "equation",
|
| 337 |
+
"img_path": "images/f712498d8f2e22eb14cf4c0d8413f3467a0c4cff08961f4addd5958711453005.jpg",
|
| 338 |
+
"text": "$$\nm _ { j , k } ^ { ( i ) } \\in \\{ 0 , 1 \\} , \\quad \\sum _ { k = 1 } ^ { C } m _ { j , k } ^ { ( i ) } = K \\forall j \\in \\{ 1 , \\ldots , C \\} \\mathrm { ~ a n d } \\forall i \\in \\{ 1 , \\ldots , L \\} .\n$$",
|
| 339 |
+
"text_format": "latex",
|
| 340 |
+
"bbox": [
|
| 341 |
+
243,
|
| 342 |
+
125,
|
| 343 |
+
718,
|
| 344 |
+
170
|
| 345 |
+
],
|
| 346 |
+
"page_idx": 4
|
| 347 |
+
},
|
| 348 |
+
{
|
| 349 |
+
"type": "text",
|
| 350 |
+
"text": "These constraints imply that each residual block receives input from exactly $K$ branches of the previous module. ",
|
| 351 |
+
"bbox": [
|
| 352 |
+
173,
|
| 353 |
+
175,
|
| 354 |
+
823,
|
| 355 |
+
203
|
| 356 |
+
],
|
| 357 |
+
"page_idx": 4
|
| 358 |
+
},
|
| 359 |
+
{
|
| 360 |
+
"type": "text",
|
| 361 |
+
"text": "Forward Propagation. During the forreal-valued branch masks for each block (i) (i) (i) $j$ ard propagation, our algorithmto sum up to 1, i.e., such that $\\begin{array} { r } { \\sum _ { k = 1 } ^ { C } \\tilde { m } _ { j , k } ^ { ( i ) } = 1 } \\end{array}$ es the . This $C$ done so that Mult( ˜m j,1, m˜ j,2, . . , m˜ j,C ) defines a proper multinomial distribution over the $C$ branch connections feeding into block $j$ . Then, the binary branch mask $\\mathbf { m } _ { j } ^ { ( i ) }$ is stochastically generated by drawing $K$ distinct samples $a _ { 1 } , a _ { 2 } , \\ldots , a _ { K } \\in \\{ 1 , \\ldots , C \\}$ from the multinomial distribution over the branch connections. Finally, the entries corresponding to the $K$ samples are activated in the binary branch mask vector, i.e., m(i)j,ak $\\dot { m } _ { j , a _ { k } } ^ { ( i ) } \\gets 1$ , for $k = 1 , . . . , K$ . The input activation volume to the residual block sampl $j$ is then computed according to Eq. 2 from the sampled binary branch masks.g from the Multinomial distribution ensures that the connections with largest ote that thevalues will $\\tilde { m } _ { j , k } ^ { ( i ) }$ be more likely to be chosen, while at the same time the stochasticity of this process allows different connectivities to be explored, particularly during early stages of the learning when the real-valued masks have still fairly uniform values. ",
|
| 362 |
+
"bbox": [
|
| 363 |
+
173,
|
| 364 |
+
218,
|
| 365 |
+
825,
|
| 366 |
+
386
|
| 367 |
+
],
|
| 368 |
+
"page_idx": 4
|
| 369 |
+
},
|
| 370 |
+
{
|
| 371 |
+
"type": "text",
|
| 372 |
+
"text": "Backward Propagation. In the backward propagation steto each branch output is obtained via back-propagation from adient and t $\\partial \\ell / \\partial y _ { k } ^ { ( i - 1 ) }$ withasks ect. $\\partial \\ell / \\partial x _ { j } ^ { \\overline { { ( i ) } } }$ $m _ { j , k } ^ { ( i ) }$ ",
|
| 373 |
+
"bbox": [
|
| 374 |
+
174,
|
| 375 |
+
401,
|
| 376 |
+
823,
|
| 377 |
+
430
|
| 378 |
+
],
|
| 379 |
+
"page_idx": 4
|
| 380 |
+
},
|
| 381 |
+
{
|
| 382 |
+
"type": "text",
|
| 383 |
+
"text": "Mask Update. In the parameter update step our algorithm computes the gradient with respect to the binary branch masks for each branch. Then, using these computed gradients and the given learning rate, it updates the real-valued branch masks via gradient descent. At this time we clip the updated real-valued branch masks to constrain them to remain within the valid interval $[ 0 , 1 ]$ . The same clipping strategy was adopted for the binary weights in the work of Courbariaux et al. (2015). ",
|
| 384 |
+
"bbox": [
|
| 385 |
+
174,
|
| 386 |
+
445,
|
| 387 |
+
825,
|
| 388 |
+
515
|
| 389 |
+
],
|
| 390 |
+
"page_idx": 4
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"type": "text",
|
| 394 |
+
"text": "As discussed in the supplementary material, after joint training over $\\theta$ and $\\mathbf { m }$ , we have found beneficial to fine-tune the weights $\\theta$ of the network with fixed binary masks (connectivity), by setting as active connections for each block $j$ in module $i$ those corresponding to the $K$ largest values in $\\tilde { \\mathbf { m } } _ { j } ^ { ( i ) }$ . Pseudocode for our training procedure is given in the supplementary material. ",
|
| 395 |
+
"bbox": [
|
| 396 |
+
174,
|
| 397 |
+
522,
|
| 398 |
+
825,
|
| 399 |
+
579
|
| 400 |
+
],
|
| 401 |
+
"page_idx": 4
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"type": "text",
|
| 405 |
+
"text": "3 EXPERIMENTS ",
|
| 406 |
+
"text_level": 1,
|
| 407 |
+
"bbox": [
|
| 408 |
+
176,
|
| 409 |
+
598,
|
| 410 |
+
326,
|
| 411 |
+
614
|
| 412 |
+
],
|
| 413 |
+
"page_idx": 4
|
| 414 |
+
},
|
| 415 |
+
{
|
| 416 |
+
"type": "text",
|
| 417 |
+
"text": "We tested our approach on the task of image categorization using several benchmarks: CIFAR10 (Krizhesvsky, 2009), CIFAR-100 (Krizhesvsky, 2009), Mini-ImageNet (Vinyals et al., 2016), as well as the full ImageNet (Deng et al., 2009). In this section we discuss results achieved on CIFAR-100 and ImageNet (Deng et al., 2009), while the results for CIFAR-10 (Krizhesvsky, 2009) and Mini-ImageNet (Vinyals et al., 2016) can be found in the Appendix. ",
|
| 418 |
+
"bbox": [
|
| 419 |
+
174,
|
| 420 |
+
630,
|
| 421 |
+
826,
|
| 422 |
+
700
|
| 423 |
+
],
|
| 424 |
+
"page_idx": 4
|
| 425 |
+
},
|
| 426 |
+
{
|
| 427 |
+
"type": "text",
|
| 428 |
+
"text": "3.1 CIFAR-100 ",
|
| 429 |
+
"text_level": 1,
|
| 430 |
+
"bbox": [
|
| 431 |
+
174,
|
| 432 |
+
717,
|
| 433 |
+
299,
|
| 434 |
+
732
|
| 435 |
+
],
|
| 436 |
+
"page_idx": 4
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"type": "text",
|
| 440 |
+
"text": "CIFAR-100 is a dataset of color images of size $3 2 \\mathbf { x } 3 2$ . It consists of 50,000 training images and 10,000 test images. Each image in CIFAR-100 is categorized into one of 100 possible classes. ",
|
| 441 |
+
"bbox": [
|
| 442 |
+
174,
|
| 443 |
+
743,
|
| 444 |
+
823,
|
| 445 |
+
772
|
| 446 |
+
],
|
| 447 |
+
"page_idx": 4
|
| 448 |
+
},
|
| 449 |
+
{
|
| 450 |
+
"type": "text",
|
| 451 |
+
"text": "Effect of fan-in $( K )$ . We start by studying the effect of the fan-in hyperparameter $( K )$ on the performance of models built and trained using our proposed approach. The fan-in defines the number of active branches feeding each residual block. For this experiment we use a model obtained by stacking $L = 6$ multi-branch residual modules, each having cardinality $C = 8$ (number of branches in each module). We use residual blocks consisting of 3 convolutional layers with a bottleneck implementing dimensionality reduction on the number of feature channels, as shown in Figure 1. The bottleneck for this experiment was set to $w = 4$ . Since each residual block consists of 3 layers, the total depth of the network in terms of learnable layers is $D = 2 + 3 L = 2 0$ . ",
|
| 452 |
+
"bbox": [
|
| 453 |
+
174,
|
| 454 |
+
776,
|
| 455 |
+
825,
|
| 456 |
+
888
|
| 457 |
+
],
|
| 458 |
+
"page_idx": 4
|
| 459 |
+
},
|
| 460 |
+
{
|
| 461 |
+
"type": "text",
|
| 462 |
+
"text": "We trained and tested this architecture using different fan-in values: $K = 1 , . . , 8$ . Note that varying $K$ does not affect the number of parameters. Thus, all these models have the same learning capacity. ",
|
| 463 |
+
"bbox": [
|
| 464 |
+
174,
|
| 465 |
+
895,
|
| 466 |
+
825,
|
| 467 |
+
924
|
| 468 |
+
],
|
| 469 |
+
"page_idx": 4
|
| 470 |
+
},
|
| 471 |
+
{
|
| 472 |
+
"type": "image",
|
| 473 |
+
"img_path": "images/e6750866c35d084002c458a652bb9670a9e538019e5d1e4238072b4a1ea295c6.jpg",
|
| 474 |
+
"image_caption": [
|
| 475 |
+
"Figure 2: Varying the fan-in $( K )$ of our model, i.e., the number of active branches provided as input to each residual block. The plot reports accuracy achieved on CIFAR-100 using a network stack of $L = 6$ ResNeXt modules having cardinality $C \\ = \\ 8$ and bottleneck width $w = 4$ . All models have the same number of parameters (0.28M). The best accuracy is obtained for $K = 4$ . "
|
| 476 |
+
],
|
| 477 |
+
"image_footnote": [],
|
| 478 |
+
"bbox": [
|
| 479 |
+
186,
|
| 480 |
+
118,
|
| 481 |
+
423,
|
| 482 |
+
256
|
| 483 |
+
],
|
| 484 |
+
"page_idx": 5
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"type": "image",
|
| 488 |
+
"img_path": "images/6d0c0f6a7a60a522b9e661bf60b1803fc9a785485c7219003df485da152b3fcb.jpg",
|
| 489 |
+
"image_caption": [
|
| 490 |
+
"Figure 3: A visualization of the fixed branch connectivity of ResNext (left) versus the connectivity learned by our method (right) using $K = 1$ ). Each green square is a residual block, each row of $C \\ = \\ 8$ square is a multibranch module. The network consists of a stack of $L = 9$ modules. Arrows indicate pathways connecting residual blocks of adjacent modules. In each net, the top red circle is a convolutional layer, the bottom circle is the final fully-connected layer. It can be noticed that MASKCONNECT learns sparse connections. The squares without $_ \\mathrm { i n / o u t }$ edges are those deemed superfluous by our algorithm and can be pruned at the end of learning. This gives rise to a branching factor that varies along the depth of the net. "
|
| 491 |
+
],
|
| 492 |
+
"image_footnote": [],
|
| 493 |
+
"bbox": [
|
| 494 |
+
470,
|
| 495 |
+
102,
|
| 496 |
+
797,
|
| 497 |
+
255
|
| 498 |
+
],
|
| 499 |
+
"page_idx": 5
|
| 500 |
+
},
|
| 501 |
+
{
|
| 502 |
+
"type": "text",
|
| 503 |
+
"text": "The results are shown in Figure 2. We can see that the best accuracy is achieved by connecting each residual block to $K = 4$ branches out of the total $C = 8$ in each module. Using a very low or very high fan-in yields lower accuracy. Note that when setting $K = C$ , there is no need to learn the masks. In this case each mask is simply replaced by an element-wise addition of the outputs from all the branches. This renders the model equivalent to ResNeXt (Xie et al., 2017), which has fixed connectivity. Based on the results of Figure 2, in all our experiments below we use $K = 4$ , since it gives the best accuracy, but also $K = 1$ , since it gives high sparsity which, as we will see shortly, implies savings in number of parameters. ",
|
| 504 |
+
"bbox": [
|
| 505 |
+
174,
|
| 506 |
+
396,
|
| 507 |
+
825,
|
| 508 |
+
507
|
| 509 |
+
],
|
| 510 |
+
"page_idx": 5
|
| 511 |
+
},
|
| 512 |
+
{
|
| 513 |
+
"type": "text",
|
| 514 |
+
"text": "Varying the architectures. In Table 1 we show the classification accuracy achieved with different architectures (the details of each architecture are listed in the Appendix). For each architecture we report results obtained using MASKCONNECT with fan-in $K = 1$ and $K = 4$ . We also include the accuracy achieved with full (as opposed to learned) connectivity, which corresponds to ResNeXt. These results show that learning the connectivity produces consistently higher accuracy than using fixed connectivity, with accuracy gains of up $2 . 2 \\%$ compared to the state-of-the-art ResNeXt model. ",
|
| 515 |
+
"bbox": [
|
| 516 |
+
174,
|
| 517 |
+
513,
|
| 518 |
+
825,
|
| 519 |
+
598
|
| 520 |
+
],
|
| 521 |
+
"page_idx": 5
|
| 522 |
+
},
|
| 523 |
+
{
|
| 524 |
+
"type": "text",
|
| 525 |
+
"text": "We note that these improvements in accuracy come at little computational training cost: the average training time overhead for learning masks and weights is about $3 9 \\%$ using our unoptimized implementation compared to learning only the weights given a fixed connectivity. Additionally, for each architecture we include models trained using sparse random connectivity (Fixed-Random). For these models, each mask is set to have $K = 4$ randomly-chosen active connections, and the connectivity is kept fixed during learning of the parameters. We can notice that the accuracy of these nets is considerably lower compared to our models, despite having the same connectivity density $X = 4$ ). This shows that the improvements of our approach over ResNeXt are not due to sparser connectivity but they are rather due to learned connectivity. ",
|
| 526 |
+
"bbox": [
|
| 527 |
+
174,
|
| 528 |
+
604,
|
| 529 |
+
825,
|
| 530 |
+
729
|
| 531 |
+
],
|
| 532 |
+
"page_idx": 5
|
| 533 |
+
},
|
| 534 |
+
{
|
| 535 |
+
"type": "text",
|
| 536 |
+
"text": "Parameter savings. Our proposed approach provides the benefit of automatically identifying during training residual blocks that are unnecessary. At the end of the training, the unused residual blocks can be pruned away. This yields savings in the number of parameters to store and in testtime computation. In Table 1, columns Train and Test under Params show the original number of parameters (used during training) and the number of parameters after pruning (used at test-time). Note that for the biggest architecture, our approach using $K = 1$ yields a parameter saving of $40 \\%$ compared to ResNeXt with full connectivity (20.5M vs 34.4M), while achieving the same accuracy. Thus, in summary, using fan-in $K = 4$ gives models that have the same number of parameters as ResNeXt but they yield higher accuracy; using fan-in $K = 1$ gives a significant saving in number of parameters and accuracy on par with ResNeXt. ",
|
| 537 |
+
"bbox": [
|
| 538 |
+
173,
|
| 539 |
+
736,
|
| 540 |
+
826,
|
| 541 |
+
876
|
| 542 |
+
],
|
| 543 |
+
"page_idx": 5
|
| 544 |
+
},
|
| 545 |
+
{
|
| 546 |
+
"type": "text",
|
| 547 |
+
"text": "Model with real-valued masks. We have also attempted to learn our models using real-valued masks by computing tensors in the forward and backward propagation with respect to masks $\\tilde { \\mathbf { m } } _ { j } ^ { ( i ) } \\in [ \\tilde { 0 , 1 } ] ^ { C }$ rather than the binary vectors $\\mathbf { m } _ { j } ^ { ( i ) } \\in \\{ 0 , 1 \\} ^ { C }$ . However, we found this variant to ",
|
| 548 |
+
"bbox": [
|
| 549 |
+
176,
|
| 550 |
+
882,
|
| 551 |
+
823,
|
| 552 |
+
925
|
| 553 |
+
],
|
| 554 |
+
"page_idx": 5
|
| 555 |
+
},
|
| 556 |
+
{
|
| 557 |
+
"type": "text",
|
| 558 |
+
"text": "Table 1: CIFAR-100 accuracies (single crop) achieved by different architectures trained using the predefined full connectivity of ResNeXt (Fixed-Full) versus the connectivity learned by our algorithm (Learned). We also include models trained using random, fixed connectivity (Fixed-Random) defined by setting $K = 4$ random active connections per branch. Each model was trained 4 times, using different random initializations. For each model we report the best test performance as well as the mean test performance computed from the 4 runs. For our method, we report performance using $K = 1$ as well as $K = 4$ . We also list the number of parameters used during training (Params-Train) and the number of parameters obtained after pruning the unused blocks (Params-Test). Our learned connectivity using $K = 4$ produces accuracy gains of up $2 . 2 \\%$ compared to the strong ResNeXt model, while using $K = 1$ yields results equivalent to ResNeXt but it induces a significant reduction in number of parameters at test time (a saving of $40 \\%$ for model $\\{ 2 9 , 6 4 , 8 \\} ,$ ). ",
|
| 559 |
+
"bbox": [
|
| 560 |
+
173,
|
| 561 |
+
99,
|
| 562 |
+
825,
|
| 563 |
+
253
|
| 564 |
+
],
|
| 565 |
+
"page_idx": 6
|
| 566 |
+
},
|
| 567 |
+
{
|
| 568 |
+
"type": "table",
|
| 569 |
+
"img_path": "images/5f27e21181ab1ff3cdf7702e2c5b3d2d4b6bccf32371107652753ea8858022cc.jpg",
|
| 570 |
+
"table_caption": [],
|
| 571 |
+
"table_footnote": [],
|
| 572 |
+
"table_body": "<table><tr><td rowspan=\"2\">Architecture {Depth (D), Bottleneck width (w),Cardinality (C)}</td><td rowspan=\"2\">Connectivity</td><td colspan=\"2\">Params</td><td>Accuracy (%)</td></tr><tr><td>Train</td><td>Test</td><td>Top-1 best (mean±std)</td></tr><tr><td rowspan=\"4\">{29,8,8}</td><td>Fixed-Full, K=8 (Xie et al.,2017)</td><td>0.86M</td><td>0.86M</td><td>73.52 (73.37±0.13)</td></tr><tr><td>Learned, K=1</td><td>0.86M</td><td>0.65M</td><td>73.91 (73.76±0.14)</td></tr><tr><td>Learned, K=4</td><td>0.86M</td><td>0.81M</td><td>75.89 (75.77±0.12)</td></tr><tr><td>Fixed-Random, K=4</td><td>0.86M</td><td>0.85M</td><td>72.85 (72.66±0.24)</td></tr><tr><td rowspan=\"4\">{29,64,8}</td><td>Fixed-Full, K=8 (Xie et al.,2017)</td><td>34.4M</td><td>34.4M</td><td>82.23 (82.12±0.12)</td></tr><tr><td>Learned, K=1</td><td>34.4M</td><td>20.5M</td><td>82.31 (82.15±0.15)</td></tr><tr><td>Learned, K=4</td><td>34.4M</td><td>32.1M</td><td>84.05 (83.94±0.11)</td></tr><tr><td>Fixed-Random, K=4</td><td>34.4M</td><td>34.3M</td><td>81.96 (81.73±0.20)</td></tr></table>",
|
| 573 |
+
"bbox": [
|
| 574 |
+
186,
|
| 575 |
+
265,
|
| 576 |
+
812,
|
| 577 |
+
433
|
| 578 |
+
],
|
| 579 |
+
"page_idx": 6
|
| 580 |
+
},
|
| 581 |
+
{
|
| 582 |
+
"type": "text",
|
| 583 |
+
"text": "yield consistently lower results compared to our models using binary masks. For example, for model $\\{ D = 2 9 , \\stackrel { \\cdot } { w } = 8 , C = 8 \\}$ the best accuracy achieved with real-valued masks is $1 . 9 3 \\%$ worse compared to that obtained with binary masks. In particular we observed that for this variant, the real-valued masks change little over training even when using large learning rates. Conversely, performing the forward and backward propagation using stochastically-sampled binary masks yields a larger exploration of connectivities and results in bigger changes of the auxiliary real-valued masks leading to better connectivity learning. ",
|
| 584 |
+
"bbox": [
|
| 585 |
+
174,
|
| 586 |
+
459,
|
| 587 |
+
825,
|
| 588 |
+
558
|
| 589 |
+
],
|
| 590 |
+
"page_idx": 6
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"type": "text",
|
| 594 |
+
"text": "Visualization of the learned connectivity. Figure 3 provides an illustration of the connectivity learned by MASKCONNECT for $K = 1$ versus the fixed connectivity of ResNeXt for model $\\{ D =$ 2 $9 , w = 8 , C = 8 \\}$ . While ResNeXt feeds the same input to all blocks of a module, our algorithm learns different input pathways for each block and yields a branching factor that varies along depth. ",
|
| 595 |
+
"bbox": [
|
| 596 |
+
174,
|
| 597 |
+
561,
|
| 598 |
+
825,
|
| 599 |
+
618
|
| 600 |
+
],
|
| 601 |
+
"page_idx": 6
|
| 602 |
+
},
|
| 603 |
+
{
|
| 604 |
+
"type": "text",
|
| 605 |
+
"text": "3.2 IMAGENET ",
|
| 606 |
+
"text_level": 1,
|
| 607 |
+
"bbox": [
|
| 608 |
+
174,
|
| 609 |
+
636,
|
| 610 |
+
292,
|
| 611 |
+
650
|
| 612 |
+
],
|
| 613 |
+
"page_idx": 6
|
| 614 |
+
},
|
| 615 |
+
{
|
| 616 |
+
"type": "text",
|
| 617 |
+
"text": "Finally, we evaluate our approach on the large-scale ImageNet 2012 dataset (Deng et al., 2009), which includes images of 1000 classes. We train our approach on the training set (1.28M images) and evaluate it on the validation set (50K images). In Table 2, we report the Top-1 and Top-5 accuracies for three different architectures. For these experiments we set $\\bar { K } = C / 2$ . We can observe that for all three architectures, our learned connectivity yields an improvement in accuracy over the fixed connectivity of ResNeXt (Xie et al., 2017). ",
|
| 618 |
+
"bbox": [
|
| 619 |
+
174,
|
| 620 |
+
661,
|
| 621 |
+
825,
|
| 622 |
+
746
|
| 623 |
+
],
|
| 624 |
+
"page_idx": 6
|
| 625 |
+
},
|
| 626 |
+
{
|
| 627 |
+
"type": "text",
|
| 628 |
+
"text": "3.3 CIFAR-10 & MINI-IMAGENET ",
|
| 629 |
+
"text_level": 1,
|
| 630 |
+
"bbox": [
|
| 631 |
+
176,
|
| 632 |
+
763,
|
| 633 |
+
429,
|
| 634 |
+
777
|
| 635 |
+
],
|
| 636 |
+
"page_idx": 6
|
| 637 |
+
},
|
| 638 |
+
{
|
| 639 |
+
"type": "text",
|
| 640 |
+
"text": "We invite the reader to review results achieved on CIFAR-10 & Mini-ImageNet in the Appendix. Also on these datasets our algorithm consistently outperforms the ResNeXt models based on fixed connectivity, with accuracy gains of up to $3 . 8 \\%$ . ",
|
| 641 |
+
"bbox": [
|
| 642 |
+
176,
|
| 643 |
+
790,
|
| 644 |
+
823,
|
| 645 |
+
832
|
| 646 |
+
],
|
| 647 |
+
"page_idx": 6
|
| 648 |
+
},
|
| 649 |
+
{
|
| 650 |
+
"type": "text",
|
| 651 |
+
"text": "4 RELATED WORK ",
|
| 652 |
+
"text_level": 1,
|
| 653 |
+
"bbox": [
|
| 654 |
+
176,
|
| 655 |
+
849,
|
| 656 |
+
344,
|
| 657 |
+
866
|
| 658 |
+
],
|
| 659 |
+
"page_idx": 6
|
| 660 |
+
},
|
| 661 |
+
{
|
| 662 |
+
"type": "text",
|
| 663 |
+
"text": "Despite their wide adoption, deep networks often require laborious model search in order to yield good results. As a result, significant research effort has been devoted to the design of algorithms for automatic model selection. However, most of this prior work falls within the genre of hyperparameter optimization (Bergstra & Bengio, 2012; Snoek et al., 2012; 2015) rather than architecture or connectivity learning. Evolutionary search has been proposed as an interesting framework to learn both the structure as well as the connections in a neural network (Wierstra et al., 2005; Floreano et al., 2008; Real et al., 2017). Architecture search has also been recently formulated as a reinforcement learning problem with impressive results (Zoph & Le, 2017). Unlike these approaches, our method is limited to learning the connectivity within a predefined architecture but it does so efficiently by gradient descent optimization of the learning objective as opposed to more costly procedures such as evolutionary search or reinforcement learning. Several authors have proposed learning connectivity by pruning unimportant weights from the network (LeCun et al., 1989; Han et al., 2015a;b; Guo et al., 2016; Han et al., 2016). However, these prior methods operate in stages where initially the network with full connectivity is learned and then connections are greedily removed according to an importance criterion. In PathNet (Fernando et al., 2017), the connectivity within a given architecture was searched via evolution. Compare to these prior approaches, our work provides the advantage of learning the connectivity by direct global optimization of the loss function of the problem at hand rather than by greedy optimization of a proxy criterion or by evolution. Our technical approach shares similarities with the “Shake-Shake” regularization recently introduced in unpublished work (Gastaldi, 2017). This procedure was demonstrated on two-branch ResNeXt models and consists in randomly scaling tensors produced by parallel branches during each training iteration while at test time the network uses uniform weighting of tensors. Conversely, our algorithm learns an optimal binary scaling of the parallel tensors with respect to the training objective and uses the resulting network with sparse connectivity at test time. Our work is also related to approaches that learn a hierarchical structure in the last one or two layers of a network in order to obtain distinct features for different categories (Murdock et al., 2016; Ahmed & Torresani, 2017). Differently from these methods, our algorithm learns efficiently connections at all depths in the network, thus optimizing over a much larger family of connectivity models. While our algorithm is limited to optimizing the connectivity structure within a predefined architecture, Adams et al. (Adams et al., 2010) proposed a nonparametric Bayesian approach that searches over an infinite network using MCMC. Saxena and Verbeek (Saxena & Verbeek, 2016) introduced convolutional neural fabric which are learnable 3D trellises that locally connect response maps at different layers of a CNN. Similarly to our work, they enable optimization over an exponentially large family of connectivities, albeit different from those considered here. ",
|
| 664 |
+
"bbox": [
|
| 665 |
+
176,
|
| 666 |
+
882,
|
| 667 |
+
825,
|
| 668 |
+
924
|
| 669 |
+
],
|
| 670 |
+
"page_idx": 6
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"type": "table",
|
| 674 |
+
"img_path": "images/60a7e4c88b37469bb4bb9d77f57006b134c38a95a65d2da33a40258ab820e044.jpg",
|
| 675 |
+
"table_caption": [
|
| 676 |
+
"Table 2: ImageNet accuracies (single crop) achieved by different architectures using the predefined connectivity of ResNeXt (Fixed-Full) versus the connectivity learned by our algorithm (Learned). "
|
| 677 |
+
],
|
| 678 |
+
"table_footnote": [],
|
| 679 |
+
"table_body": "<table><tr><td>Architecture</td><td>Connectivity</td><td colspan=\"2\">Accuracy</td></tr><tr><td>{Depth (D), Bottleneck width (w), Cardinality (C}</td><td></td><td>Top-1</td><td>Top-5</td></tr><tr><td rowspan=\"2\">{50,4,32}</td><td>Fixed-Full, K=32 (Xie et al., 2017)</td><td>77.8</td><td>93.3</td></tr><tr><td>Learned, K=16</td><td>79.1</td><td>94.1</td></tr><tr><td rowspan=\"2\">{101,4,32}</td><td>Fixed-Full, K=32 (Xie et al.,2017)</td><td>78.8</td><td>94.1</td></tr><tr><td>Learned, K=16</td><td>79.5</td><td>94.5</td></tr><tr><td rowspan=\"2\">{101,4,64}</td><td>Fixed-Full, K=64 (Xie et al., 2017)</td><td>79.6</td><td>94.7</td></tr><tr><td>Learned, K=32</td><td>79.8</td><td>94.8</td></tr></table>",
|
| 680 |
+
"bbox": [
|
| 681 |
+
215,
|
| 682 |
+
141,
|
| 683 |
+
781,
|
| 684 |
+
289
|
| 685 |
+
],
|
| 686 |
+
"page_idx": 7
|
| 687 |
+
},
|
| 688 |
+
{
|
| 689 |
+
"type": "text",
|
| 690 |
+
"text": "",
|
| 691 |
+
"bbox": [
|
| 692 |
+
174,
|
| 693 |
+
314,
|
| 694 |
+
825,
|
| 695 |
+
729
|
| 696 |
+
],
|
| 697 |
+
"page_idx": 7
|
| 698 |
+
},
|
| 699 |
+
{
|
| 700 |
+
"type": "text",
|
| 701 |
+
"text": "5 CONCLUSIONS ",
|
| 702 |
+
"text_level": 1,
|
| 703 |
+
"bbox": [
|
| 704 |
+
176,
|
| 705 |
+
746,
|
| 706 |
+
328,
|
| 707 |
+
761
|
| 708 |
+
],
|
| 709 |
+
"page_idx": 7
|
| 710 |
+
},
|
| 711 |
+
{
|
| 712 |
+
"type": "text",
|
| 713 |
+
"text": "In this paper we introduced an algorithm to learn the connectivity of deep multi-branch networks. The problem is formulated as a single joint optimization over the weights and the branch connections of the model. We tested our approach on challenging image categorization benchmarks where it led to significant accuracy improvements over the state-of-the-art ResNeXt model. An added benefit of our approach is that it can automatically identify superfluous blocks, which can be pruned without impact on accuracy for more efficient testing and for reducing the number of parameters to store. ",
|
| 714 |
+
"bbox": [
|
| 715 |
+
174,
|
| 716 |
+
776,
|
| 717 |
+
825,
|
| 718 |
+
861
|
| 719 |
+
],
|
| 720 |
+
"page_idx": 7
|
| 721 |
+
},
|
| 722 |
+
{
|
| 723 |
+
"type": "text",
|
| 724 |
+
"text": "While our experiments were focused on a particular multi-branch architecture (ResNeXt) and a specific form of building block (residual block), we expect the benefits of our approach to extend to other modules and network structures. For example, it could be applied to learn the connectivity of skip-connections in DenseNets (Huang et al., 2017), which are currently based on predefined connectivity rules. In this paper, our masks perform non-parametric additive aggregation of the branch outputs. It would be interesting to experiment with learnable (parametric) aggregations of the outputs from the individual branches. Our approach is limited to learning connectivity within a given, fixed architecture. Future work will explore the use of learnable masks for architecture discovery. ",
|
| 725 |
+
"bbox": [
|
| 726 |
+
174,
|
| 727 |
+
867,
|
| 728 |
+
823,
|
| 729 |
+
924
|
| 730 |
+
],
|
| 731 |
+
"page_idx": 7
|
| 732 |
+
},
|
| 733 |
+
{
|
| 734 |
+
"type": "text",
|
| 735 |
+
"text": "",
|
| 736 |
+
"bbox": [
|
| 737 |
+
174,
|
| 738 |
+
103,
|
| 739 |
+
825,
|
| 740 |
+
159
|
| 741 |
+
],
|
| 742 |
+
"page_idx": 8
|
| 743 |
+
},
|
| 744 |
+
{
|
| 745 |
+
"type": "text",
|
| 746 |
+
"text": "REFERENCES ",
|
| 747 |
+
"text_level": 1,
|
| 748 |
+
"bbox": [
|
| 749 |
+
174,
|
| 750 |
+
180,
|
| 751 |
+
285,
|
| 752 |
+
195
|
| 753 |
+
],
|
| 754 |
+
"page_idx": 8
|
| 755 |
+
},
|
| 756 |
+
{
|
| 757 |
+
"type": "text",
|
| 758 |
+
"text": "Ryan Prescott Adams, Hanna M. Wallach, and Zoubin Ghahramani. Learning the structure of deep sparse graphical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, AISTATS 2010, Chia Laguna Resort, Sardinia, Italy, May 13-15, 2010, pp. 1–8, 2010. ",
|
| 759 |
+
"bbox": [
|
| 760 |
+
174,
|
| 761 |
+
203,
|
| 762 |
+
826,
|
| 763 |
+
260
|
| 764 |
+
],
|
| 765 |
+
"page_idx": 8
|
| 766 |
+
},
|
| 767 |
+
{
|
| 768 |
+
"type": "text",
|
| 769 |
+
"text": "Karim Ahmed and Lorenzo Torresani. Branchconnect: Large-scale visual recognition with learned branch connections. CoRR, abs/1704.06010, 2017. URL http://arxiv.org/abs/1704. 06010. ",
|
| 770 |
+
"bbox": [
|
| 771 |
+
174,
|
| 772 |
+
268,
|
| 773 |
+
825,
|
| 774 |
+
310
|
| 775 |
+
],
|
| 776 |
+
"page_idx": 8
|
| 777 |
+
},
|
| 778 |
+
{
|
| 779 |
+
"type": "text",
|
| 780 |
+
"text": "James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13:281–305, 2012. ",
|
| 781 |
+
"bbox": [
|
| 782 |
+
173,
|
| 783 |
+
320,
|
| 784 |
+
823,
|
| 785 |
+
348
|
| 786 |
+
],
|
| 787 |
+
"page_idx": 8
|
| 788 |
+
},
|
| 789 |
+
{
|
| 790 |
+
"type": "text",
|
| 791 |
+
"text": "Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in Neural Information Processing Systems 28, Montreal, Quebec, Canada, pp. 3123–3131, 2015. ",
|
| 792 |
+
"bbox": [
|
| 793 |
+
173,
|
| 794 |
+
357,
|
| 795 |
+
825,
|
| 796 |
+
400
|
| 797 |
+
],
|
| 798 |
+
"page_idx": 8
|
| 799 |
+
},
|
| 800 |
+
{
|
| 801 |
+
"type": "text",
|
| 802 |
+
"text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pp. 248–255, 2009. ",
|
| 803 |
+
"bbox": [
|
| 804 |
+
173,
|
| 805 |
+
409,
|
| 806 |
+
825,
|
| 807 |
+
452
|
| 808 |
+
],
|
| 809 |
+
"page_idx": 8
|
| 810 |
+
},
|
| 811 |
+
{
|
| 812 |
+
"type": "text",
|
| 813 |
+
"text": "Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu, Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. CoRR, abs/1701.08734, 2017. URL http://arxiv.org/abs/1701.08734. ",
|
| 814 |
+
"bbox": [
|
| 815 |
+
174,
|
| 816 |
+
460,
|
| 817 |
+
825,
|
| 818 |
+
503
|
| 819 |
+
],
|
| 820 |
+
"page_idx": 8
|
| 821 |
+
},
|
| 822 |
+
{
|
| 823 |
+
"type": "text",
|
| 824 |
+
"text": "Dario Floreano, Peter Durr, and Claudio Mattiussi. Neuroevolution: from architectures to learning. ¨ Evolutionary Intelligence, 1(1):47–62, 2008. ",
|
| 825 |
+
"bbox": [
|
| 826 |
+
174,
|
| 827 |
+
512,
|
| 828 |
+
823,
|
| 829 |
+
541
|
| 830 |
+
],
|
| 831 |
+
"page_idx": 8
|
| 832 |
+
},
|
| 833 |
+
{
|
| 834 |
+
"type": "text",
|
| 835 |
+
"text": "Xavier Gastaldi. Shake-shake regularization. CoRR, abs/1705.07485, 2017. URL http://arxiv. org/abs/1705.07485. ",
|
| 836 |
+
"bbox": [
|
| 837 |
+
173,
|
| 838 |
+
549,
|
| 839 |
+
825,
|
| 840 |
+
578
|
| 841 |
+
],
|
| 842 |
+
"page_idx": 8
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"type": "text",
|
| 846 |
+
"text": "Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, AISTATS 2011, Fort Lauderdale, USA, April 11-13, 2011, pp. 315–323, 2011. ",
|
| 847 |
+
"bbox": [
|
| 848 |
+
173,
|
| 849 |
+
587,
|
| 850 |
+
823,
|
| 851 |
+
630
|
| 852 |
+
],
|
| 853 |
+
"page_idx": 8
|
| 854 |
+
},
|
| 855 |
+
{
|
| 856 |
+
"type": "text",
|
| 857 |
+
"text": "Yiwen Guo, Anbang Yao, and Yurong Chen. Dynamic network surgery for efficient dnns. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pp. 1379–1387, 2016. ",
|
| 858 |
+
"bbox": [
|
| 859 |
+
176,
|
| 860 |
+
637,
|
| 861 |
+
825,
|
| 862 |
+
681
|
| 863 |
+
],
|
| 864 |
+
"page_idx": 8
|
| 865 |
+
},
|
| 866 |
+
{
|
| 867 |
+
"type": "text",
|
| 868 |
+
"text": "Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding. In International Conference on Learning Representations (ICLR), 2015a. ",
|
| 869 |
+
"bbox": [
|
| 870 |
+
174,
|
| 871 |
+
689,
|
| 872 |
+
823,
|
| 873 |
+
732
|
| 874 |
+
],
|
| 875 |
+
"page_idx": 8
|
| 876 |
+
},
|
| 877 |
+
{
|
| 878 |
+
"type": "text",
|
| 879 |
+
"text": "Song Han, Jeff Pool, John Tran, and William J. Dally. Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems 28, Montreal, Quebec, Canada, pp. 1135–1143, 2015b. ",
|
| 880 |
+
"bbox": [
|
| 881 |
+
174,
|
| 882 |
+
741,
|
| 883 |
+
825,
|
| 884 |
+
784
|
| 885 |
+
],
|
| 886 |
+
"page_idx": 8
|
| 887 |
+
},
|
| 888 |
+
{
|
| 889 |
+
"type": "text",
|
| 890 |
+
"text": "Song Han, Jeff Pool, Sharan Narang, Huizi Mao, Shijian Tang, Erich Elsen, Bryan Catanzaro, John Tran, and William J. Dally. DSD: regularizing deep neural networks with dense-sparse-dense training flow. In International Conference on Learning Representations (ICLR), 2016. ",
|
| 891 |
+
"bbox": [
|
| 892 |
+
176,
|
| 893 |
+
791,
|
| 894 |
+
823,
|
| 895 |
+
835
|
| 896 |
+
],
|
| 897 |
+
"page_idx": 8
|
| 898 |
+
},
|
| 899 |
+
{
|
| 900 |
+
"type": "text",
|
| 901 |
+
"text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), 2016 IEEE Conference on, 2016. ",
|
| 902 |
+
"bbox": [
|
| 903 |
+
176,
|
| 904 |
+
843,
|
| 905 |
+
823,
|
| 906 |
+
886
|
| 907 |
+
],
|
| 908 |
+
"page_idx": 8
|
| 909 |
+
},
|
| 910 |
+
{
|
| 911 |
+
"type": "text",
|
| 912 |
+
"text": "Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, 2017. ",
|
| 913 |
+
"bbox": [
|
| 914 |
+
174,
|
| 915 |
+
895,
|
| 916 |
+
821,
|
| 917 |
+
924
|
| 918 |
+
],
|
| 919 |
+
"page_idx": 8
|
| 920 |
+
},
|
| 921 |
+
{
|
| 922 |
+
"type": "text",
|
| 923 |
+
"text": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pp. 448–456, 2015. ",
|
| 924 |
+
"bbox": [
|
| 925 |
+
176,
|
| 926 |
+
103,
|
| 927 |
+
821,
|
| 928 |
+
146
|
| 929 |
+
],
|
| 930 |
+
"page_idx": 9
|
| 931 |
+
},
|
| 932 |
+
{
|
| 933 |
+
"type": "text",
|
| 934 |
+
"text": "Alex Krizhesvsky. Learning multiple layers of features from tiny images, 2009. Technical Report https://www.cs.toronto.edu/˜kriz/learning-features-2009-TR.pdf. ",
|
| 935 |
+
"bbox": [
|
| 936 |
+
173,
|
| 937 |
+
155,
|
| 938 |
+
823,
|
| 939 |
+
184
|
| 940 |
+
],
|
| 941 |
+
"page_idx": 9
|
| 942 |
+
},
|
| 943 |
+
{
|
| 944 |
+
"type": "text",
|
| 945 |
+
"text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25, Lake Tahoe, Nevada, United States., pp. 1106–1114, 2012. ",
|
| 946 |
+
"bbox": [
|
| 947 |
+
176,
|
| 948 |
+
191,
|
| 949 |
+
823,
|
| 950 |
+
234
|
| 951 |
+
],
|
| 952 |
+
"page_idx": 9
|
| 953 |
+
},
|
| 954 |
+
{
|
| 955 |
+
"type": "text",
|
| 956 |
+
"text": "Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-30, 1989], pp. 598–605, 1989. ",
|
| 957 |
+
"bbox": [
|
| 958 |
+
174,
|
| 959 |
+
242,
|
| 960 |
+
825,
|
| 961 |
+
286
|
| 962 |
+
],
|
| 963 |
+
"page_idx": 9
|
| 964 |
+
},
|
| 965 |
+
{
|
| 966 |
+
"type": "text",
|
| 967 |
+
"text": "Andrew L Maas, Awni Y Hannun, and Andrew Y Ng. Rectifier nonlinearities improve neural network acoustic models. Proc. ICML, 30:1, 2013. ",
|
| 968 |
+
"bbox": [
|
| 969 |
+
171,
|
| 970 |
+
294,
|
| 971 |
+
823,
|
| 972 |
+
324
|
| 973 |
+
],
|
| 974 |
+
"page_idx": 9
|
| 975 |
+
},
|
| 976 |
+
{
|
| 977 |
+
"type": "text",
|
| 978 |
+
"text": "Calvin Murdock, Zhen Li, Howard Zhou, and Tom Duerig. Blockout: Dynamic model selection for hierarchical deep networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pp. 2583–2591, 2016. ",
|
| 979 |
+
"bbox": [
|
| 980 |
+
176,
|
| 981 |
+
332,
|
| 982 |
+
825,
|
| 983 |
+
375
|
| 984 |
+
],
|
| 985 |
+
"page_idx": 9
|
| 986 |
+
},
|
| 987 |
+
{
|
| 988 |
+
"type": "text",
|
| 989 |
+
"text": "Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR), 2017. ",
|
| 990 |
+
"bbox": [
|
| 991 |
+
174,
|
| 992 |
+
382,
|
| 993 |
+
821,
|
| 994 |
+
412
|
| 995 |
+
],
|
| 996 |
+
"page_idx": 9
|
| 997 |
+
},
|
| 998 |
+
{
|
| 999 |
+
"type": "text",
|
| 1000 |
+
"text": "Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Quoc V. Le, and Alex Kurakin. Large-scale evolution of image classifiers. CoRR, abs/1703.01041, 2017. ",
|
| 1001 |
+
"bbox": [
|
| 1002 |
+
173,
|
| 1003 |
+
420,
|
| 1004 |
+
825,
|
| 1005 |
+
449
|
| 1006 |
+
],
|
| 1007 |
+
"page_idx": 9
|
| 1008 |
+
},
|
| 1009 |
+
{
|
| 1010 |
+
"type": "text",
|
| 1011 |
+
"text": "Shreyas Saxena and Jakob Verbeek. Convolutional neural fabrics. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pp. 4053–4061, 2016. ",
|
| 1012 |
+
"bbox": [
|
| 1013 |
+
176,
|
| 1014 |
+
457,
|
| 1015 |
+
825,
|
| 1016 |
+
501
|
| 1017 |
+
],
|
| 1018 |
+
"page_idx": 9
|
| 1019 |
+
},
|
| 1020 |
+
{
|
| 1021 |
+
"type": "text",
|
| 1022 |
+
"text": "Pierre Sermanet, David Eigen, Xiang Zhang, Michael Mathieu, Rob Fergus, and Yann LeCun. ¨ Overfeat: Integrated recognition, localization and detection using convolutional networks. In International Conference on Learning Representations (ICLR), 2013. ",
|
| 1023 |
+
"bbox": [
|
| 1024 |
+
178,
|
| 1025 |
+
508,
|
| 1026 |
+
823,
|
| 1027 |
+
551
|
| 1028 |
+
],
|
| 1029 |
+
"page_idx": 9
|
| 1030 |
+
},
|
| 1031 |
+
{
|
| 1032 |
+
"type": "text",
|
| 1033 |
+
"text": "Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015. ",
|
| 1034 |
+
"bbox": [
|
| 1035 |
+
173,
|
| 1036 |
+
559,
|
| 1037 |
+
825,
|
| 1038 |
+
589
|
| 1039 |
+
],
|
| 1040 |
+
"page_idx": 9
|
| 1041 |
+
},
|
| 1042 |
+
{
|
| 1043 |
+
"type": "text",
|
| 1044 |
+
"text": "Jasper Snoek, Hugo Larochelle, and Ryan P. Adams. Practical bayesian optimization of machine learning algorithms. In Advances in Neural Information Processing Systems 25, Lake Tahoe, Nevada, United States., pp. 2960–2968, 2012. ",
|
| 1045 |
+
"bbox": [
|
| 1046 |
+
174,
|
| 1047 |
+
597,
|
| 1048 |
+
823,
|
| 1049 |
+
641
|
| 1050 |
+
],
|
| 1051 |
+
"page_idx": 9
|
| 1052 |
+
},
|
| 1053 |
+
{
|
| 1054 |
+
"type": "text",
|
| 1055 |
+
"text": "Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan Kiros, Nadathur Satish, Narayanan Sundaram, Md. Mostofa Ali Patwary, Prabhat, and Ryan P. Adams. Scalable bayesian optimization using deep neural networks. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pp. 2171–2180, 2015. ",
|
| 1056 |
+
"bbox": [
|
| 1057 |
+
173,
|
| 1058 |
+
648,
|
| 1059 |
+
826,
|
| 1060 |
+
705
|
| 1061 |
+
],
|
| 1062 |
+
"page_idx": 9
|
| 1063 |
+
},
|
| 1064 |
+
{
|
| 1065 |
+
"type": "text",
|
| 1066 |
+
"text": "Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pp. 1–9, 2015. ",
|
| 1067 |
+
"bbox": [
|
| 1068 |
+
173,
|
| 1069 |
+
713,
|
| 1070 |
+
826,
|
| 1071 |
+
770
|
| 1072 |
+
],
|
| 1073 |
+
"page_idx": 9
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"type": "text",
|
| 1077 |
+
"text": "Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems 29, Barcelona, Spain, pp. 3630–3638, 2016. ",
|
| 1078 |
+
"bbox": [
|
| 1079 |
+
174,
|
| 1080 |
+
779,
|
| 1081 |
+
825,
|
| 1082 |
+
821
|
| 1083 |
+
],
|
| 1084 |
+
"page_idx": 9
|
| 1085 |
+
},
|
| 1086 |
+
{
|
| 1087 |
+
"type": "text",
|
| 1088 |
+
"text": "Daan Wierstra, Faustino J. Gomez, and Jurgen Schmidhuber. Modeling systems with internal state ¨ using evolino. In Genetic and Evolutionary Computation Conference, GECCO 2005, Proceedings, Washington DC, USA, June 25-29, 2005, pp. 1795–1802, 2005. ",
|
| 1089 |
+
"bbox": [
|
| 1090 |
+
173,
|
| 1091 |
+
830,
|
| 1092 |
+
823,
|
| 1093 |
+
873
|
| 1094 |
+
],
|
| 1095 |
+
"page_idx": 9
|
| 1096 |
+
},
|
| 1097 |
+
{
|
| 1098 |
+
"type": "text",
|
| 1099 |
+
"text": "Saining Xie, Ross B. Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual ´ transformations for deep neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, 2017. ",
|
| 1100 |
+
"bbox": [
|
| 1101 |
+
174,
|
| 1102 |
+
882,
|
| 1103 |
+
825,
|
| 1104 |
+
924
|
| 1105 |
+
],
|
| 1106 |
+
"page_idx": 9
|
| 1107 |
+
},
|
| 1108 |
+
{
|
| 1109 |
+
"type": "text",
|
| 1110 |
+
"text": "Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I, pp. 818–833, 2014. ",
|
| 1111 |
+
"bbox": [
|
| 1112 |
+
176,
|
| 1113 |
+
103,
|
| 1114 |
+
825,
|
| 1115 |
+
146
|
| 1116 |
+
],
|
| 1117 |
+
"page_idx": 10
|
| 1118 |
+
},
|
| 1119 |
+
{
|
| 1120 |
+
"type": "text",
|
| 1121 |
+
"text": "Barret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning. In International Conference on Learning Representations (ICLR), 2017. ",
|
| 1122 |
+
"bbox": [
|
| 1123 |
+
173,
|
| 1124 |
+
155,
|
| 1125 |
+
823,
|
| 1126 |
+
184
|
| 1127 |
+
],
|
| 1128 |
+
"page_idx": 10
|
| 1129 |
+
},
|
| 1130 |
+
{
|
| 1131 |
+
"type": "text",
|
| 1132 |
+
"text": "A PSEUDOCODE OF THE ALGORITHM ",
|
| 1133 |
+
"text_level": 1,
|
| 1134 |
+
"bbox": [
|
| 1135 |
+
176,
|
| 1136 |
+
102,
|
| 1137 |
+
496,
|
| 1138 |
+
118
|
| 1139 |
+
],
|
| 1140 |
+
"page_idx": 11
|
| 1141 |
+
},
|
| 1142 |
+
{
|
| 1143 |
+
"type": "text",
|
| 1144 |
+
"text": "Algorithm 1 MASKCONNECT training algorithm. ",
|
| 1145 |
+
"text_level": 1,
|
| 1146 |
+
"bbox": [
|
| 1147 |
+
174,
|
| 1148 |
+
148,
|
| 1149 |
+
491,
|
| 1150 |
+
164
|
| 1151 |
+
],
|
| 1152 |
+
"page_idx": 11
|
| 1153 |
+
},
|
| 1154 |
+
{
|
| 1155 |
+
"type": "text",
|
| 1156 |
+
"text": "Input: a minibatch of labeled examples $( x ^ { i } , y ^ { i } )$ , $C$ : cardinality (number of branches), $K$ : fan-in (number of active branch connections), $\\eta$ : learning rate, $\\ell$ : the loss over the minibatch, $\\tilde { \\mathbf { m } } _ { j } ^ { ( i ) } \\in [ 0 , 1 ] ^ { C }$ : real-valued branch masks for block $j$ in module $_ { i }$ from previous training iteration. \nOutput: updated $\\tilde { \\mathbf { m } } _ { j } ^ { ( i ) }$ \n1. Forward Propagation: \nNormalize the real-valued mask to sum up to 1: m˜ (i)j,k $\\begin{array} { r } { \\tilde { m } _ { j , k } ^ { ( i ) } \\gets \\frac { \\tilde { m } _ { j , k } ^ { ( i ) } } { \\sum _ { k ^ { \\prime } = 1 } ^ { C } \\tilde { m } _ { j , k ^ { \\prime } } ^ { ( i ) } } } \\end{array}$ , for $j = 1 , \\ldots , C$ \nReset binary mask: $\\mathbf { m } _ { j } ^ { ( i ) } \\gets \\mathbf { 0 }$ \nDraw $K$ distinct samples from multinomial mask distribution: \n$a _ { 1 } , a _ { 2 } , \\ldots , a _ { K } \\gets \\hat { \\mathbf { M u l t } } ( \\tilde { m } _ { j , 1 } ^ { ( i ) } , \\tilde { m } _ { j , 2 } ^ { ( i ) } , \\ldots , \\tilde { m } _ { j , C } ^ { ( i ) } )$ \nSet active binary mask based on drawn samples: \n$m _ { j , a _ { k } } ^ { ( i ) } \\gets 1$ for $k = 1 , . . . , K$ \nCompute output $\\mathbf { x } _ { j } ^ { ( i ) }$ of the mask, given branch activations $\\begin{array} { r } { \\mathbf { y } _ { k } ^ { ( i - 1 ) } \\colon \\mathbf { x } _ { j } ^ { ( i ) } \\sum _ { k = 1 } ^ { C } m _ { j , k } ^ { ( i ) } \\cdot \\mathbf { y } _ { k } ^ { ( i - 1 ) } } \\end{array}$ \nj2. Backward Propagation: \nCompute Compute $\\frac { \\partial \\ell } { \\partial \\mathbf { x } _ { j } ^ { ( i ) } }$ om fro ∂ y(i)j \\` \n$\\frac { \\partial \\ell } { \\partial \\mathbf { y } _ { k } ^ { ( i - 1 ) } }$ ∂x(i)j and m(i)j,k \n3. Parameter Update: \nCompute m˜ (i) ← c $\\frac { \\partial \\ell } { \\partial m _ { j , k } ^ { ( i ) } }$ given (i) − η ∂ x(i)j d y(ik an\\` ) $\\mathbf { y } _ { k } ^ { ( i - 1 ) }$ \n$\\begin{array} { r } { \\tilde { m } _ { j , k } ^ { ( i ) } \\mathrm { c l i p } ( \\tilde { m } _ { j , k } ^ { ( i ) } - \\eta \\cdot \\frac { \\partial \\ell } { \\partial m _ { j , k } ^ { ( i ) } } ) } \\end{array}$ ",
|
| 1157 |
+
"bbox": [
|
| 1158 |
+
184,
|
| 1159 |
+
167,
|
| 1160 |
+
821,
|
| 1161 |
+
479
|
| 1162 |
+
],
|
| 1163 |
+
"page_idx": 11
|
| 1164 |
+
},
|
| 1165 |
+
{
|
| 1166 |
+
"type": "text",
|
| 1167 |
+
"text": "B EXPERIMENTS ON CIFAR-10 ",
|
| 1168 |
+
"text_level": 1,
|
| 1169 |
+
"bbox": [
|
| 1170 |
+
176,
|
| 1171 |
+
522,
|
| 1172 |
+
454,
|
| 1173 |
+
539
|
| 1174 |
+
],
|
| 1175 |
+
"page_idx": 11
|
| 1176 |
+
},
|
| 1177 |
+
{
|
| 1178 |
+
"type": "text",
|
| 1179 |
+
"text": "The CIFAR-10 dataset consists of color images of size $3 2 \\mathrm { x } 3 2$ . The training set contains 50,000 images, the testing set 10,000 images. Each image in CIFAR-10 is categorized into one of 10 possible classes. In Table 3, we report the performance of different models trained on CIFAR-10. From these results we can observe that our models using learned connectivity achieve consistently better performance over the equivalent models trained with the fixed connectivity (Xie et al., 2017). ",
|
| 1180 |
+
"bbox": [
|
| 1181 |
+
173,
|
| 1182 |
+
558,
|
| 1183 |
+
825,
|
| 1184 |
+
628
|
| 1185 |
+
],
|
| 1186 |
+
"page_idx": 11
|
| 1187 |
+
},
|
| 1188 |
+
{
|
| 1189 |
+
"type": "table",
|
| 1190 |
+
"img_path": "images/e74a8db7e8801b16efd544cd12d3ccda73c62da9aa84832ef60672f9e1504699.jpg",
|
| 1191 |
+
"table_caption": [
|
| 1192 |
+
"Table 3: CIFAR-10 accuracies (single crop) achieved by different multi-branch architectures trained using the predefined connectivity of ResNeXt (Fixed-Full) versus the connectivity learned by our algorithm (Learned). Each model was trained 4 times, using different random initializations. For each model we report the best test performance as well as the mean test performance computed from the 4 runs. "
|
| 1193 |
+
],
|
| 1194 |
+
"table_footnote": [],
|
| 1195 |
+
"table_body": "<table><tr><td>Architecture</td><td>Connectivity</td><td>Accuracy (%)</td></tr><tr><td>{Depth (D), Bottleneck width (w), Cardinality (C)}</td><td></td><td>Top-1 best (mean±std)</td></tr><tr><td>{20,4,8}</td><td>Fixed-Full K=8 (Xie et al., 2017) Learned K=4</td><td>91.39 (91.13±0.11) 92.85 (92.76±0.10)</td></tr><tr><td>{29,4,8}</td><td>Fixed-Full K=8 (Xie et al.,2017) Learned K=4</td><td>92.77 (92.65±0.09) 93.88 (93.76±0.12)</td></tr><tr><td>{29,8.8}</td><td>Fixed-Full K=8 (Xie et al.,2017) Learned K=4</td><td>93.26 (93.14±0.11) 95.11 (94.96±0.12)</td></tr><tr><td>{29,64,8}</td><td>Fixed-Full K=8 (Xie et al.,2017) Learned K=4</td><td>96.35 (96.23±0.12) 96.83 (96.73±0.11)</td></tr></table>",
|
| 1196 |
+
"bbox": [
|
| 1197 |
+
228,
|
| 1198 |
+
728,
|
| 1199 |
+
769,
|
| 1200 |
+
909
|
| 1201 |
+
],
|
| 1202 |
+
"page_idx": 11
|
| 1203 |
+
},
|
| 1204 |
+
{
|
| 1205 |
+
"type": "table",
|
| 1206 |
+
"img_path": "images/ffc6bc89b29486b686e1d7a1d95d68cbd0b093026f2be48949a1f8d7bd51fb54.jpg",
|
| 1207 |
+
"table_caption": [
|
| 1208 |
+
"Table 4: Mini-ImageNet accuracies achieved by different multi-branch networks trained using the predefined full connectivity of ResNeXt (Fixed-Full) versus the connectivity learned by our algorithm (Learned). Additionally, we include models trained using random fixed connectivity (Fixed-Random) for $K = 4$ . For each model we report the best and the mean test performance computed from 4 different training runs. Our method for joint learning of weights and connectivity yields a gain of over $3 \\%$ in Top-1 accuracy over ResNeXt, which uses the same architectures but a fixed branch connectivity. "
|
| 1209 |
+
],
|
| 1210 |
+
"table_footnote": [],
|
| 1211 |
+
"table_body": "<table><tr><td>Architecture</td><td>Connectivity</td><td>Accuracy</td></tr><tr><td>{Depth (D), Bottleneck width (w), Cardinality (C}</td><td></td><td>Top-1 best (mean±std)</td></tr><tr><td rowspan=\"3\">{20,4,8}</td><td>Fixed-Full K=8 (Xie et al., 2017)</td><td>62.12 (61.86±0.15)</td></tr><tr><td>Learned K=4</td><td>66.09 (65.94±0.16)</td></tr><tr><td>Fixed-Random K=4</td><td>62.42 (61.81±0.32)</td></tr><tr><td rowspan=\"3\">{29,8,8}</td><td>Fixed-Full K=8 (Xie et al., 2017)</td><td>68.11 (67.89±0.19)</td></tr><tr><td>Learned K=4</td><td>71.36 (71.18±0.19)</td></tr><tr><td>Fixed-Random K=4</td><td>67.97 (67.53±0.20)</td></tr></table>",
|
| 1212 |
+
"bbox": [
|
| 1213 |
+
228,
|
| 1214 |
+
210,
|
| 1215 |
+
769,
|
| 1216 |
+
353
|
| 1217 |
+
],
|
| 1218 |
+
"page_idx": 12
|
| 1219 |
+
},
|
| 1220 |
+
{
|
| 1221 |
+
"type": "text",
|
| 1222 |
+
"text": "C EXPERIMENTS ON MINI-IMAGENET ",
|
| 1223 |
+
"text_level": 1,
|
| 1224 |
+
"bbox": [
|
| 1225 |
+
174,
|
| 1226 |
+
397,
|
| 1227 |
+
508,
|
| 1228 |
+
414
|
| 1229 |
+
],
|
| 1230 |
+
"page_idx": 12
|
| 1231 |
+
},
|
| 1232 |
+
{
|
| 1233 |
+
"type": "text",
|
| 1234 |
+
"text": "Mini-ImageNet is a subset of the full ImageNet (Deng et al., 2009) dataset. It was used in (Vinyals et al., 2016; Ravi & Larochelle, 2017). It is created by randomly selecting 100 classes from the full ImageNet (Deng et al., 2009). For each class, 600 images are randomly selected. We use 500 examples per class for training, and the other 100 examples per class for testing. The selected images are resized to size 84x84 pixels as in (Vinyals et al., 2016; Ravi & Larochelle, 2017). The advantage of this dataset is that it poses the recognition challenges typical of the ImageNet photos but at the same time it does not need require the powerful resources needed to train on the full ImageNet dataset. This allows to include the additional baselines involving random fixed connectivity (Fixed-Random). ",
|
| 1235 |
+
"bbox": [
|
| 1236 |
+
173,
|
| 1237 |
+
443,
|
| 1238 |
+
825,
|
| 1239 |
+
555
|
| 1240 |
+
],
|
| 1241 |
+
"page_idx": 12
|
| 1242 |
+
},
|
| 1243 |
+
{
|
| 1244 |
+
"type": "text",
|
| 1245 |
+
"text": "We report the performance of different models trained on Mini-ImageNet in Table 4. From these results, we see that our models using learned connectivity with fan-in $K { = } 4$ yield a nice accuracy gain over the same models trained with the fixed full connectivity of ResNeXt (Xie et al., 2017). The absolute improvement (in Top-1 accuracy) is $3 . 8 7 \\%$ for the 20-layer network and $3 . 1 7 \\%$ for the 29-layer network. We can notice that the accuracy of the models with fixed random connectivity (Fixed-Random) is considerably lower compared to our nets with learned connectivity, despite having the same connectivity density $K = 4$ ). This shows that the improvement of our approach over ResNeXt is not due to sparser connectivity but it is rather due to learned connectivity. ",
|
| 1246 |
+
"bbox": [
|
| 1247 |
+
174,
|
| 1248 |
+
561,
|
| 1249 |
+
825,
|
| 1250 |
+
672
|
| 1251 |
+
],
|
| 1252 |
+
"page_idx": 12
|
| 1253 |
+
},
|
| 1254 |
+
{
|
| 1255 |
+
"type": "text",
|
| 1256 |
+
"text": "D VISUALIZATIONS OF LEARNED CONNECTIVITY ",
|
| 1257 |
+
"text_level": 1,
|
| 1258 |
+
"bbox": [
|
| 1259 |
+
173,
|
| 1260 |
+
718,
|
| 1261 |
+
596,
|
| 1262 |
+
733
|
| 1263 |
+
],
|
| 1264 |
+
"page_idx": 12
|
| 1265 |
+
},
|
| 1266 |
+
{
|
| 1267 |
+
"type": "text",
|
| 1268 |
+
"text": "The plot in Figure 4 shows how the number of active branches varies as a function of the module depth for model $\\{ D = 2 9 , w = 4 , C = 8 \\}$ trained on CIFAR-100. For $K = 1$ , we can observe that the number of active branches tends to be larger for deep modules (closer to the output layer) compared to early modules (closer to the input). We observed this phenomenon consistently for all architectures. This suggests that having many parallel threads of computation is particularly important in deep layers of the network. Conversely, the setting $K = 4$ tends to produce a fairly uniform number of active branches across the modules and the number is quite close to the maximum value $C$ . For this reason, there is little saving in terms of number of parameters when using $K = 4$ , as there are rarely unused blocks. ",
|
| 1269 |
+
"bbox": [
|
| 1270 |
+
174,
|
| 1271 |
+
762,
|
| 1272 |
+
825,
|
| 1273 |
+
888
|
| 1274 |
+
],
|
| 1275 |
+
"page_idx": 12
|
| 1276 |
+
},
|
| 1277 |
+
{
|
| 1278 |
+
"type": "text",
|
| 1279 |
+
"text": "The plot in Figure 5 shows the number of active branches as a function of module depth for model $\\{ D \\stackrel { - } { = } 5 0 , w \\stackrel { - } { = } 4 , C = 3 2 \\}$ trained on ImageNet, using $K = 1 6$ . ",
|
| 1280 |
+
"bbox": [
|
| 1281 |
+
173,
|
| 1282 |
+
895,
|
| 1283 |
+
823,
|
| 1284 |
+
924
|
| 1285 |
+
],
|
| 1286 |
+
"page_idx": 12
|
| 1287 |
+
},
|
| 1288 |
+
{
|
| 1289 |
+
"type": "image",
|
| 1290 |
+
"img_path": "images/f2fe0b9ece80b9214c7f504a43cdfbeb591158a51b9131cf89c2fcf589a32129.jpg",
|
| 1291 |
+
"image_caption": [
|
| 1292 |
+
"Figure 4: Number of active branches as a function of module depth for model $\\{ D = 2 9 , w = 4 , C =$ $8 \\}$ trained on CIFAR-100. We report how the number of active branches varies for model trained with fan-in $K = 1$ as well as for the net trained with $K = 4$ . The setting $K = 1$ tends to leave many blocks unused, especially in the early modules of the network. "
|
| 1293 |
+
],
|
| 1294 |
+
"image_footnote": [],
|
| 1295 |
+
"bbox": [
|
| 1296 |
+
186,
|
| 1297 |
+
162,
|
| 1298 |
+
807,
|
| 1299 |
+
391
|
| 1300 |
+
],
|
| 1301 |
+
"page_idx": 13
|
| 1302 |
+
},
|
| 1303 |
+
{
|
| 1304 |
+
"type": "image",
|
| 1305 |
+
"img_path": "images/1f86c77fd1a50dce61f4b663010ecd737e0fb6a98e59de51a355f0f0c5abcb25.jpg",
|
| 1306 |
+
"image_caption": [
|
| 1307 |
+
"Figure 5: Number of active branches as a function of module depth for model $\\{ D = 5 0 , w = 4 , C =$ $\\mathrm { 3 2 } \\bar \\}$ trained on ImageNet, using fan-in $K = 1 6$ . "
|
| 1308 |
+
],
|
| 1309 |
+
"image_footnote": [],
|
| 1310 |
+
"bbox": [
|
| 1311 |
+
186,
|
| 1312 |
+
592,
|
| 1313 |
+
808,
|
| 1314 |
+
819
|
| 1315 |
+
],
|
| 1316 |
+
"page_idx": 13
|
| 1317 |
+
},
|
| 1318 |
+
{
|
| 1319 |
+
"type": "text",
|
| 1320 |
+
"text": "Table 5: Specifications of the architectures used in our experiments on the CIFAR-10 and CIFAR-100 datasets. The architectures differ in terms of depth $( D )$ , bottleneck width $( w )$ , and cardinality $( C )$ . Inside the brackets we specify the residual block used in each multi-branch module by listing the number of input channels, the size of the convolutional filters, as well as the number of filters (number of output channels). To the right of each bracket we list the cardinality (i.e., the number of parallel branches in the module). $\\times 2$ means that the same multi-branch module is stacked twice. The first layer for all models is a convolutional layer with 16 filters of size $3 \\times 3$ . The last layer performs global average pooling followed by a softmax. ",
|
| 1321 |
+
"bbox": [
|
| 1322 |
+
173,
|
| 1323 |
+
99,
|
| 1324 |
+
825,
|
| 1325 |
+
212
|
| 1326 |
+
],
|
| 1327 |
+
"page_idx": 14
|
| 1328 |
+
},
|
| 1329 |
+
{
|
| 1330 |
+
"type": "table",
|
| 1331 |
+
"img_path": "images/9b07d16c1f30eba3f39fd005121f306c11c5ec8cafeb4a5f36a06bc2bb391406.jpg",
|
| 1332 |
+
"table_caption": [],
|
| 1333 |
+
"table_footnote": [],
|
| 1334 |
+
"table_body": "<table><tr><td colspan=\"3\">{D=20,w=4,C=8}</td><td colspan=\"3\">{D=29,w=4,C=8}</td><td colspan=\"3\">{D=29,w=8,C=8}</td><td colspan=\"2\">{D=29,w=64,C=8}</td></tr><tr><td colspan=\"3\">3,3×3,16</td><td colspan=\"2\">3,3×3,16</td><td colspan=\"3\">3,3×3,16</td><td colspan=\"3\">3,3×3,64</td></tr><tr><td>16,1×1,4 4,3×3,4 [4,1×1,64 [64,1×1,4] 4,3×3,4</td><td>(C=8) (C=8)</td><td></td><td>[16,1×1,4 4,3×3,4 [4,1×1,64 [64,1×1,4]</td><td>(C=8)</td><td>16,1×1,8 8,3×3,8 8,1×1,64 [64,1×1,8]</td><td>(C=8)</td><td></td><td>64,1×1,64 64,3×3,64 64,1×1,256 [256,1×1,64</td><td>(C=8)</td></tr><tr><td>[4,1×1,64 [64,1×1,8 8,3×3,8 [8,1×1,128 [128,1×1,8 8,3×3,8</td><td></td><td>(C=8)</td><td>4,3×3,4 [4,1×1,64 [64,1×1,8 8,3×3,8 [8,1×1,128</td><td>(C=8),×2 (C=8)</td><td>8,3×3,8 [8,1×1,64] [64,1×1,16 16,3×3,16 16,1×1,128</td><td>(C=8),×2 (C=8)</td><td>64,3×3,64</td><td>64,1×1,256 256,1×1,128 128,3×3,128 128,1×1,512</td><td>(C=8),×2 (C=8)</td></tr><tr><td>8,1×1,128 128,1×1,16</td><td></td><td>(C=8)</td><td>[128,1×1,8 8,3×3,8 [8,1×1,128 128,1×1,16</td><td>(C=8),×2</td><td>[128,1×1,16 16,3×3,16 16,1×1,128 128,1×1,32</td><td>(C=8),×2</td><td></td><td>[512,1×1,128] 128,3×3,128 128,1×1,512] 512,1×1,256</td><td>(C=8), ×2</td></tr><tr><td colspan=\"2\">16,3×3,16 [16,1×1,256 [256,1×1,16] 16,3×3,16 16,1×1,256</td><td>(C=8) (C=8)</td><td colspan=\"2\">16,3×3,16 (C=8) [16,1×1,256 [256,1×1,16] 16,3×3,16 (C=8),×2</td><td>32,3×3,32 32,1×1,256 256,1×1,32 32,3×3,32</td><td>(C=8)</td><td></td><td>256,3×3,256 [256,1×1,1024 [1024,1×1,256 256,3×3,256</td><td>(C=8)</td></tr><tr><td colspan=\"2\">Average Pool 100 fc, softmax</td><td></td><td colspan=\"2\">[16,1×1,256 Average Pool 100 fc,softmax</td><td>32,1×1,256 100 fc, softmax</td><td>Average Pool</td><td>(C=8),×2</td><td colspan=\"2\">(C=8),×2 256,1×1,1024 Average Pool</td></tr></table>",
|
| 1335 |
+
"bbox": [
|
| 1336 |
+
215,
|
| 1337 |
+
224,
|
| 1338 |
+
782,
|
| 1339 |
+
508
|
| 1340 |
+
],
|
| 1341 |
+
"page_idx": 14
|
| 1342 |
+
},
|
| 1343 |
+
{
|
| 1344 |
+
"type": "text",
|
| 1345 |
+
"text": "E IMPLEMENTATION DETAILS ",
|
| 1346 |
+
"text_level": 1,
|
| 1347 |
+
"bbox": [
|
| 1348 |
+
174,
|
| 1349 |
+
537,
|
| 1350 |
+
434,
|
| 1351 |
+
554
|
| 1352 |
+
],
|
| 1353 |
+
"page_idx": 14
|
| 1354 |
+
},
|
| 1355 |
+
{
|
| 1356 |
+
"type": "text",
|
| 1357 |
+
"text": "E.1 ARCHITECTURES AND SETTINGS FOR EXPERIMENTS ON CIFAR-100 AND CIFAR-10 ",
|
| 1358 |
+
"bbox": [
|
| 1359 |
+
173,
|
| 1360 |
+
573,
|
| 1361 |
+
803,
|
| 1362 |
+
587
|
| 1363 |
+
],
|
| 1364 |
+
"page_idx": 14
|
| 1365 |
+
},
|
| 1366 |
+
{
|
| 1367 |
+
"type": "text",
|
| 1368 |
+
"text": "The specifications of the architectures used in all our experiments on CIFAR-10 and CIFAR-100 are given in Table 5. ",
|
| 1369 |
+
"bbox": [
|
| 1370 |
+
176,
|
| 1371 |
+
601,
|
| 1372 |
+
821,
|
| 1373 |
+
628
|
| 1374 |
+
],
|
| 1375 |
+
"page_idx": 14
|
| 1376 |
+
},
|
| 1377 |
+
{
|
| 1378 |
+
"type": "text",
|
| 1379 |
+
"text": "Several of these architectures are those presented in the original ResNeXt paper (Xie et al., 2017) and are trained using the same setup, including the data augmentation strategy.Four pixels are padded on each side of the input image, and a $3 2 \\mathbf { x } 3 2$ crop is randomly sampled from the padded image or its horizontal flip, with per-pixel mean subtracted (Krizhevsky et al., 2012). For testing, we use the original $3 2 \\mathbf { x } 3 2$ image. The stacks have output feature map of size 32, 16, and 8 respectively. The models are trained on 8 GPUs with a mini-batch size of 128 (16 per GPU), with a weight decay of 0.0005 and momentum of 0.9. We adopt four incremental training phases with a total of 320 epochs. In phase 1 we train the model for 120 epochs with a learning rate of 0.1 for the convolutional and fully-connected layers, and a learning rate of 0.2 for the masks. In phase 2 we freeze the connectivity by setting as active connections for each block those corresponding to its top- $K$ values in the masks. With these fixed learned connectivity, we finetune the model from phase $^ { l }$ for 100 epochs with a learning rate of 0.1 for the weights. Then, in phase 3 we finetune the weights of the model from phase 2 for 50 epochs with a learning rate of 0.01 using again the fixed learned connectivity from phase 1. Finally, in phase 4 we finetune the weights of the model from phase 3 for 50 epochs with a learning rate of 0.001. ",
|
| 1380 |
+
"bbox": [
|
| 1381 |
+
174,
|
| 1382 |
+
636,
|
| 1383 |
+
825,
|
| 1384 |
+
844
|
| 1385 |
+
],
|
| 1386 |
+
"page_idx": 14
|
| 1387 |
+
},
|
| 1388 |
+
{
|
| 1389 |
+
"type": "text",
|
| 1390 |
+
"text": "E.2 ARCHITECTURES AND SETTINGS FOR EXPERIMENTS ON IMAGENET ",
|
| 1391 |
+
"text_level": 1,
|
| 1392 |
+
"bbox": [
|
| 1393 |
+
176,
|
| 1394 |
+
868,
|
| 1395 |
+
683,
|
| 1396 |
+
882
|
| 1397 |
+
],
|
| 1398 |
+
"page_idx": 14
|
| 1399 |
+
},
|
| 1400 |
+
{
|
| 1401 |
+
"type": "text",
|
| 1402 |
+
"text": "The architectures for our ImageNet experiments are those specified in the original ResNeXt paper (Xie et al., 2017). ",
|
| 1403 |
+
"bbox": [
|
| 1404 |
+
173,
|
| 1405 |
+
895,
|
| 1406 |
+
820,
|
| 1407 |
+
924
|
| 1408 |
+
],
|
| 1409 |
+
"page_idx": 14
|
| 1410 |
+
},
|
| 1411 |
+
{
|
| 1412 |
+
"type": "text",
|
| 1413 |
+
"text": "Table 6: Mini-ImageNet architectures with varying depth $( D )$ , and bottleneck width $( w )$ . Inside the brackets we specify the residual block used in each multi-branch module by listing the number of input channels, the size of the convolutional filters, as well as the number of filters (number of output channels). To the right of each bracket we list the cardinality $( C )$ (i.e., the number of parallel branches in the module). $\\times 2$ means that the same multi-branch module is stacked twice. ",
|
| 1414 |
+
"bbox": [
|
| 1415 |
+
173,
|
| 1416 |
+
101,
|
| 1417 |
+
825,
|
| 1418 |
+
170
|
| 1419 |
+
],
|
| 1420 |
+
"page_idx": 15
|
| 1421 |
+
},
|
| 1422 |
+
{
|
| 1423 |
+
"type": "table",
|
| 1424 |
+
"img_path": "images/0a8e28c8c1530f4badc22fc607550ed33294d75d1b72f406e8728ea96221e5bb.jpg",
|
| 1425 |
+
"table_caption": [],
|
| 1426 |
+
"table_footnote": [],
|
| 1427 |
+
"table_body": "<table><tr><td colspan=\"2\">{D=20,w=4,C=8}</td><td colspan=\"2\">{D=29,=8,C=8}</td></tr><tr><td colspan=\"2\">3,3×3,16</td><td colspan=\"2\">3,3×3,16</td></tr><tr><td colspan=\"2\">Max Pool,3×3,stride=2</td><td colspan=\"2\">Max Pool,3×3,stride=2</td></tr><tr><td colspan=\"2\">[16,1×1,4 4,3×3,4 (C=8) [4,1×1,64 [64,1×1,4 4,3×3,4 (C=8) [4,1×1,64</td><td colspan=\"2\">[16,1×1,8 8,3×3,8 (C=8) [8,1×1,64 [64,1×1,8] 8,3×3,8 (C=8),×2</td></tr><tr><td colspan=\"2\">[64,1×1,8 8,3×3,8 [8,1×1,128 [128,1×1,8 8,3×3,8</td><td colspan=\"2\">[64,1×1,16 (C=8) 16,3×3,16 16,1×1,128 [128,1×1,16 (C=8) 16,3×3,16</td></tr><tr><td colspan=\"2\">[8,1×1,128 [128,1×1,16 16,3×3,16 (C=8)</td><td colspan=\"2\">(C=8),×2 [16,1×1,128] 128,1×1,32</td></tr><tr><td colspan=\"2\">16,1×1,256 [256,1×1,16] 16,3×3,16 (C=8) 16,1×1,256</td><td colspan=\"2\">[32,1×1,256 [256,1×1,32] 32,3×3,32 (C=8),×2 32,1×1,256]</td></tr><tr><td colspan=\"2\">Average Pool 100 fc, softmax</td><td colspan=\"2\">Average Pool 100 fc,softmax</td></tr></table>",
|
| 1428 |
+
"bbox": [
|
| 1429 |
+
362,
|
| 1430 |
+
181,
|
| 1431 |
+
633,
|
| 1432 |
+
483
|
| 1433 |
+
],
|
| 1434 |
+
"page_idx": 15
|
| 1435 |
+
},
|
| 1436 |
+
{
|
| 1437 |
+
"type": "text",
|
| 1438 |
+
"text": "Also for these experiments, we follow the data augmentation strategy described in (Xie et al., 2017). The input image has size $2 2 4 \\mathbf { x } 2 2 4$ and it is randomly cropped from the resized original image. We use a mini-batch size of 256 on 8 GPUs (32 per GPU), with a weight decay of 0.0001 and a momentum of 0.9. We use four incremental training phases with a total of 120 epochs. In phase 1 we train the model for 30 epochs with a learning rate of 0.1 for the convolutional and fully-connected layers, and a learning rate of 0.2 for the masks. In phase 2 we finetune the model from phase $^ { l }$ for another 30 epochs with a learning rate of 0.1 and a learning rate of 0.0 for the masks (i.e., we use the fixed connectivity learned in phase 1). In phase 3 we finetune the weights from phase 2 for 30 epochs with a learning rate of 0.01 and the learning rate of the masks is 0.0. Finally, in phase 4 we finetune the weights from phase 3 for 30 epochs with a learning rate of 0.001 while the learning rate of the masks is still set to 0.0. ",
|
| 1439 |
+
"bbox": [
|
| 1440 |
+
173,
|
| 1441 |
+
507,
|
| 1442 |
+
826,
|
| 1443 |
+
660
|
| 1444 |
+
],
|
| 1445 |
+
"page_idx": 15
|
| 1446 |
+
},
|
| 1447 |
+
{
|
| 1448 |
+
"type": "text",
|
| 1449 |
+
"text": "E.3 ARCHITECTURES AND SETTINGS FOR EXPERIMENTS ON MINI-IMAGENET ",
|
| 1450 |
+
"text_level": 1,
|
| 1451 |
+
"bbox": [
|
| 1452 |
+
176,
|
| 1453 |
+
678,
|
| 1454 |
+
723,
|
| 1455 |
+
691
|
| 1456 |
+
],
|
| 1457 |
+
"page_idx": 15
|
| 1458 |
+
},
|
| 1459 |
+
{
|
| 1460 |
+
"type": "text",
|
| 1461 |
+
"text": "For the experiments on the Mini-ImageNet dataset, a 64x64 crop is randomly sampled from the scaled 84x84 image or its horizontal flip, with per-pixel mean subtracted (Krizhevsky et al., 2012). For testing, we use the center $6 4 \\mathrm { x } 6 4$ crop. The specifications of the models are identical to the CIFAR-100 models used in the previous subsection, except that the first input convolutional layer in the network is followed by a max pooling layer. The models are trained on 8 GPUs with a mini-batch size of 256 (32 per GPU), with a weight decay of 0.0005 and momentum of 0.9. Similar to training CIFAR-100 dataset, we also adopt four incremental training phases with a total of 320 epochs. ",
|
| 1462 |
+
"bbox": [
|
| 1463 |
+
174,
|
| 1464 |
+
703,
|
| 1465 |
+
825,
|
| 1466 |
+
801
|
| 1467 |
+
],
|
| 1468 |
+
"page_idx": 15
|
| 1469 |
+
}
|
| 1470 |
+
]
|
parse/train/Sy3fJXbA-/Sy3fJXbA-_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/Sy3fJXbA-/Sy3fJXbA-_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/n4IMHNb8_f/n4IMHNb8_f.md
ADDED
|
@@ -0,0 +1,374 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# DIFFERENTIABLE SPATIAL PLANNING USING TRANSFORMERS
|
| 2 |
+
|
| 3 |
+
Anonymous authors Paper under double-blind review
|
| 4 |
+
|
| 5 |
+
# ABSTRACT
|
| 6 |
+
|
| 7 |
+
We consider the problem of spatial path planning. In contrast to the classical solutions which optimize a new plan from scratch and assume access to the full map with ground truth obstacle locations, we learn a planner from the data in a differentiable manner that allows us to leverage statistical regularities from past data. We propose Spatial Planning Transformers (SPT), which given an obstacle map learns to generate actions by planning over long-range spatial dependencies, unlike prior data-driven planners that propagate information locally via convolutional structure in an iterative manner. In the setting where the ground truth map is not known to the agent, we leverage pre-trained SPTs to in an end-to-end framework that has the structure of mapper and planner built into it which allows seamless generalization to out-of-distribution maps and goals. SPTs outperform prior stateof-the-art across all the setups for both manipulation and navigation tasks, leading to an absolute improvement of $7 . 1 9 \%$ .
|
| 8 |
+
|
| 9 |
+
# 1 INTRODUCTION
|
| 10 |
+
|
| 11 |
+
The problem of path planning has been a bedrock of robotics. Given an obstacle map of an environment and a goal location in the map, the task is to output a shortest path to the goal location starting from any position in the map. We consider path planning with spatial maps. Building a top-down spatial map is common practice in robotic navigation as it provides a natural representation of physical space (Durrant-Whyte & Bailey, 2006). In fact, even Observations robotic manipulation can also be naturally phrased via spatial map using the formalism of configuration spaces (Lozano-Perez, 1990), as shown in Figure 1. This problem has been studied in robotics for several decades, and classic goto planning algorithms involve Dijkstra et al. (1959), PRM (Kavraki et al., 1996), RRT (LaValle & Kuffner Jr, 2001), RRT\* (Karaman & Frazzoli, 2011), etc.
|
| 12 |
+
|
| 13 |
+

|
| 14 |
+
Figure 1: Spatial Path Planning: The raw obserPerception Planning Map Actionsvations (top left) and obstacles can be represented spatially via top-down map in navigation (left) and 3via configuration space in manipulation (right).
|
| 15 |
+
|
| 16 |
+
Our objective is to develop methods that can learn to plan from data. However, a natural question is why do we need learning for a problem which has stable classical solutions? There are two key reasons. First, classical methods do not capture statistical regularities present in the natural world, (for e.g., walls are mostly parallel or perpendicular to each other), because they optimize a plan from scratch for each new setup. This also makes analytical planning methods to be often slow at inference time which is an issue in dynamic scenarios where a more reactive policy might be required for fast adaptation from failures. A learned planner represented via a neural network can not only capture regularities but is also efficient at inference as the plan is just a result of forward-pass through the network. Second, a critical assumption of classical algorithms is that a global ground-truth obstacle space must be known to the agent ahead of time. This is in stark contrast to biological agents where cognitive maps are not pixel-accurate ground truth location of agents, but built through actions in the environment, e.g., rats build an implicit map of the environment incrementally through trajectories enabling them to take shortcuts (Tolman, 1948). A learned solution could not only provides the ability to deal with partial, noisy maps and but also help build maps on the fly while acting in the environment by backpropagating through the generated long-range plans.
|
| 17 |
+
|
| 18 |
+

|
| 19 |
+
Figure 2: Local vs Long-distance value propagation. Figure showing an example of number of iterations required to propagate distance values over a map using local and long-distance value propagation. The obstacle map and goal location shown on the left and the distance value predictions over 5 iterations is shown on the right (distance values increase from blue to yellow). Prior methods based on convolutional networks use local value propagation and require many iterations to propagate values accurately over the whole map (top right). Our method is based on long-distance value propagation between points without any obstacle between them. This type of value propagation can cover the whole map in 3 iterations in this example (bottom right).
|
| 20 |
+
|
| 21 |
+
Several recent works have proposed data-driven path planning models (Tamar et al., 2016; Karkus et al., 2017; Nardelli et al., 2018; Lee et al., 2018). Similar to how classical algorithms, like Dijkstra et al. (1959), move outward from the goal one cell at a time to predict distances iteratively based on the obstacles in the map, current learning-based spatial planning models propagate distance values in only a local neighborhood using convolutional networks. This kind of local value propagation requires $\mathcal { O } ( M )$ iterations, where $M$ is the map dimension. In theory, however, the optimal paths can be computed much more efficiently with total iterations that are on the order of number of obstacles rather than the map size. For instance, consider two points with no obstacle between, an efficient planner could directly connect them with interpolated distance. Nonetheless, this is possible only if the model can perform long-range reasoning in the obstacle space which is a challenge.
|
| 22 |
+
|
| 23 |
+
In this work, our goal is to capture this long-range spatial relationship. Transformers (Vaswani et al., 2017) are well suited for this kind of computation as they treat the inputs as sets and propagate information across all the points within the set. Building on this, we propose Spatial Planning Transformers (SPT) which consists of attention heads that can attend to any part of the input. The key idea behind the design of the proposed model is that value can be propagated between distant points if there are no obstacles between them. This would reduce the number of required iterations to $\mathcal { O } ( n _ { O } )$ where $n _ { O }$ is the number of obstacles in the map. Figure 2 shows a simple example where long-distance value propagation can cover the entire map within 3 iterations while local value propagation takes more than 5 iterations – this difference grows with the complexity of the obstacle space and map size. We compare the performance of SPTs with prior state-of-the-art learned planning approaches, VIN (Tamar et al., 2016) and GPPN (Lee et al., 2018), across both navigation as well as manipulation setups. SPTs achieve significantly higher accuracy than these prior methods for the same inference time and show over $1 0 \%$ absolute improvement when the maps are large.
|
| 24 |
+
|
| 25 |
+
Next, we turn to the case when the map is not known apriori. This is a practical setting when the agent either has access to a partially known map or just know it through the trajectories. In psychology, this is known as going from route knowledge to survey knowledege (Golledge et al., 1995) where animals aggregate the knowledge from trajectories into a cognitive map. We operationalize this setup by formulating an end-to-end differentiable framework, which in contrast to having a generic parametric policy learning (Glasmachers, 2017), has the structure of mapper and planner built into it. We first pre-train the SPT planner to capture a generic data-driven prior, and then backpropagate through it to learn a mapper that maps raw observations to obstacle map. This allows us to learn without requiring map supervision or interaction. Learned mapper and planner not only allow us to plan for new goal locations at inference but also generalize to unseen maps.
|
| 26 |
+
|
| 27 |
+
Our end-to-end mapping and planning approach provides a unified solution for both navigation and manipulation. We perform thorough experiments in both manipulation and real-world navigation maps as well as manipulation. Our approach outperforms prior state-of-the-art by a margin on both mapping and planning accuracy without assuming access to the map at training or inference.
|
| 28 |
+
|
| 29 |
+

|
| 30 |
+
Figure 3: Spatial Planning Transformer (SPT). Figure showing an overview of the proposed Spatial Planning Transformer model. It consists of 3 modules: an Encoder $E$ to encode the input, a Transformer network $T$ responsible for planning, and a Decoder $D$ decoding the output of the Transformer into action distances.
|
| 31 |
+
|
| 32 |
+
# 2 PRELIMINARIES AND PROBLEM DEFINITION
|
| 33 |
+
|
| 34 |
+
We represent the input spatial map as a matrix, $m$ , of size $M \times M$ with each element being 1, denoting obstacles, or 0, denoting free space. The goal location is also represented as a matrix, $g$ , of size $M \times M$ with exactly one element being 1, denoting the goal location, and rest 0s. The input to the spatial planning model, $x$ , consists of matrices $m$ and $g$ stacked, $x = [ m , g ]$ , where $x$ is of size $2 \times M \times M$ . The objective of the planning model is to predict $y$ which is of size $M \times M$ , consisting of action distances of corresponding locations from the goal. Here, action distance is defined to be the minimum number of actions required to reach the goal.
|
| 35 |
+
|
| 36 |
+
For navigation, $m$ is a top-down obstacle map, and $g$ represents the goal position on this map. For manipulation, $m$ represents the obstacles in the configuration space of 2-dof planar arm with joint angles denoted by $\theta _ { 1 }$ and $\theta _ { 2 }$ . Each element $( i , j )$ in $m$ indicate whether the configuration of the arm with joint angles $\theta _ { 1 } = i$ and $\theta _ { 2 } = j$ , would lead to a collision. $g$ represents the goal configuration of the arm. In the first set of experiments, we will assume that $m$ is known and in the second set of experiments, $m$ is not known and the agent receives observations, $o$ , from its sensors instead.
|
| 37 |
+
|
| 38 |
+
# 3 METHODS
|
| 39 |
+
|
| 40 |
+
We design a spatial planning model capable of long-distance information propagation. We first describe the design of this spatial planning module, called Spatial Planning Transformer(SPT), shown in Figure 3, which takes in a map and a goal as input and predict the distance to the goal from all locations. We then describe how the SPT model can be used as a planning module to train end-to-end learning models, which take in raw sensory observations and goal location as input and predict action distances without having access to the map.
|
| 41 |
+
|
| 42 |
+
# 3.1 SPT: SPATIAL PLANNING TRANSFORMERS
|
| 43 |
+
|
| 44 |
+
To propogate information over distant points, we use the Transformer (Vaswani et al., 2017) architecture. The self-attention mechanism in a Transformer can learn to attend to any element of the input. The allows the model to learn spatial reasoning over the whole map concurrently. Figure 3 shows an overview of the SPT model, which consists of three modules, an Encoder $E$ to encode the input, a Transformer network $T$ responsible for spatial planning, and a Decoder $D$ decoding the output of the Transformer into action distances.
|
| 45 |
+
|
| 46 |
+
Encoder. The Encoder $E$ computes the encoding of the input $x$ : $x _ { I } ~ = ~ E ( x )$ . The input x ∈ {0, 1}2×M×M consisting of the map and goal is first passed through a 2-layer convolutional network (LeCun et al., 1998) with ReLU activations to compute an embedding for each input element. Both layers have a kernel size of $1 \times 1$ , which ensures that the embedding of all the obstacles is identical to each other, and the same holds true for free space and the goal location. The output of this convolutional network is of size $d \times M \times M$ , where $d$ is the embedding size. This output is then flattened to get $x _ { I }$ of size $d \times M ^ { 2 }$ and passed into the Transformer network.
|
| 47 |
+
|
| 48 |
+

|
| 49 |
+
Figure 4: End-to-end Mapping and Planning. Figure showing an overview of end-to-end mapping and planning model for both the navigation and manipulation tasks.
|
| 50 |
+
|
| 51 |
+
Transformer. The Transformer network $T$ converts the input encoding into the output encoding: $x _ { O } = T ( x _ { I } )$ . It first adds the positional encoding to the input encoding. The positional encoding enables the Transformer model to distinguish between the obstacles at different locations. We use a constant sinusoidal positional encoding (Vaswani et al., 2017):
|
| 52 |
+
|
| 53 |
+
$$
|
| 54 |
+
p _ { ( 2 i , j ) } = \sin ( j / C ^ { 2 i / d } ) , \qquad p _ { ( 2 i + 1 , j ) } = \cos ( j / C ^ { 2 i / d } )
|
| 55 |
+
$$
|
| 56 |
+
|
| 57 |
+
where $p \in \mathcal { R } ^ { d \times M ^ { 2 } }$ is the positional encoding, $j \in \{ 1 , 2 , \dots , M ^ { 2 } \}$ is the position of the input, $i \in \{ 1 , 2 , \ldots , d / 2 \}$ , and $C \bar { = } M ^ { 2 }$ is a constant.
|
| 58 |
+
|
| 59 |
+
The positional encoding of each element is added to their corresponding input encoding to get $Z = x _ { I } + p . Z$ is then passed through $N = 5$ identical Transformer layers $( f _ { \mathrm { T L } } )$ to get $x _ { O }$ .
|
| 60 |
+
|
| 61 |
+
Decoder. The Decoder $D$ computes the distance prediction $\hat { y }$ from $x _ { O }$ using a position-wise fully connected layer:
|
| 62 |
+
|
| 63 |
+
$$
|
| 64 |
+
\hat { y } _ { i } = W _ { D } ^ { T } x _ { T , i } + b _ { D }
|
| 65 |
+
$$
|
| 66 |
+
|
| 67 |
+
where $\boldsymbol { x } _ { T , i } \in \mathcal { R } ^ { d \times 1 }$ is the input at position $i \in { 1 , 2 , \dots , M ^ { 2 } }$ , $W _ { D } \in \mathcal { R } ^ { d \times 1 } , b _ { D } \in \mathcal { R }$ are parameters of the Decoder shared across all positions $i$ and $\hat { y } _ { i } \in \mathcal { R }$ is the distance prediction at position $i$ . The distance prediction at all position are reshaped into a matrix to get the final prediction $\hat { y } \in \mathcal { R } ^ { M , M }$ . The entire model is trained using pairs of input $x$ and output $y$ datapoints with mean-squared error as the loss function.
|
| 68 |
+
|
| 69 |
+
# 3.2 END-TO-END MAPPING AND PLANNING
|
| 70 |
+
|
| 71 |
+
The SPT model described above is designed to predict action distances given a map as input. However, in many applications, the map of the environment is often not known. In such cases, an autonomous agent working in a realistic environment needs to predict the map from raw sensory observations. While it is possible to train a separate mapper model to predict maps from observations, this often requires map annotations which are expensive to obtain and often inaccurate. In contrast, demonstration trajectories consisting of observations and optimal actions are more readily available or easier to obtain in many robotics applications. One of the key benefits of learning-based differentiable spatial planning is that it can be used to learn mapping just from action supervision in an end-to-end fashion without having access to ground-truth maps. To demonstrate this benefit, we train an endto-end mapping and planning model to predict action distances from sensor observations for both navigation and manipulation tasks.
|
| 72 |
+
|
| 73 |
+
The end-to-end mapping and planning model consists of two modules, a Mapper $( f _ { M } )$ and a Planner $( f _ { P } )$ , as illustrated in Figure 4. The Mapper is used to predict the map $\hat { m }$ from sensor observations $o$ and the Planner is a spatial planning model to predict action distances, $\hat { y }$ , from the predicted map $\hat { m }$ :
|
| 74 |
+
|
| 75 |
+
$$
|
| 76 |
+
\hat { y } = f _ { P } ( \hat { m } ) = f _ { P } ( f _ { M } ( o ) )
|
| 77 |
+
$$
|
| 78 |
+
|
| 79 |
+
For navigation, $o$ is the set of first-person RGB camera images each of size $3 \times H \times W$ . We sample 4 images, one for each orientation, at each valid location in the map. For invalid locations, we pass an empty image of 0s for all orientations. Thus, for a map of size $M \times M$ , observation $o$ consist of $4 M ^ { 2 }$ images for all locations and 4 orientations similar to the setup in Lee et al. (2018). For manipulation, $o$ is a top-down view of the operational space with obstacles of size $P \times P$ , where each element is 1 or 0 denoting obstacles or free space. We use different Mapper architectures for navigation and manipulation experiments.
|
| 80 |
+
|
| 81 |
+
The Navigation Mapper module predicts a single value between 0 and 1 for each image in $o$ indicating whether the cell in the front of the image is an obstacle or not. The architecture of the Navigation mapper consists of ResNet18 convolutional layers followed by fully-connected layers (see Appendix for details). Each cell can have upto 4 predictions (from images corresponding to the four neighboring cells facing the current cell), which are aggregated using max pooling to get a single prediction. This map prediction consisting of continuous values is passed to the Planner module.
|
| 82 |
+
|
| 83 |
+
The Manipulation Mapper module needs to predict which configurations of the arm would lead to a collision. To predict whether a particular configuration $( \theta _ { 1 } , \theta _ { 2 } )$ , the mapper module needs to check whether any point in this configuration consists of an obstacle. A Transformer-based model is well suited to learn this function as well as it can attend to arbitrary locations in the operational space to predict the obstacles in the configuration space. We use the same architecture of the SPT model as the Manipulation Mapper as well, with the only difference being the encoder consisting of $3 \times 3$ kernel size convolutional layers instead of $1 \times 1$ to encode the $P \times P$ observation space to a $M \times M$ representation.
|
| 84 |
+
|
| 85 |
+
The Planner module is the entire SPT model with encoder, transformer and decoder units as described in the previous subsection. It is pretrained on synthetic maps and its weights are frozen during end-to-end training. We train the entire end-to-end Mapping and Planning model with pairs of input observations $o$ and output action distances $y$ using standard supervised learning with the meansquared error as the loss function: $\begin{array} { r } { \mathcal { L } = M S E ( y , \hat { y } ) = 1 / | L | \sum _ { i \in L } ( y _ { i } - \hat { y } _ { i } ) ^ { 2 } } \end{array}$ , where $L$ is the set of navigable locations. Since the planning module is pretrained and it expects a structured map input, the mapper model needs to predict the map accurately such that the predicted map, when passed through the planner, minimizes the action level loss.
|
| 86 |
+
|
| 87 |
+
# 4 EXPERIMENTS & RESULTS: SPATIAL PLANNING
|
| 88 |
+
|
| 89 |
+
Datasets. We generate synthetic datasets for training the spatial planning models for both navigation and manipulation settings. For the navigation setting, we perform experiments with $M \times M$ maps with three different map sizes, $M \in \{ 1 5 , 3 0 , 5 0 \}$ . For manipulation, we experiment with two map sizes, $M \in \{ 1 8 , 3 6 \}$ , corresponding to $2 0 ^ { \circ }$ and $1 0 ^ { \circ }$ bins for each link. In each map, we randomly generate $o _ { m i n } = 0$ to $o _ { m a x } = 5$ obstacles. Dataset generation details are provided in the Appendix.
|
| 90 |
+
|
| 91 |
+
For both the settings, we generate training, validation, and test sets of size $1 0 0 K / 5 K / 5 K$ maps. The set of maps in each set are distinct. For each map, we choose a random free space cell as the goal location. The action space consists of 4 actions: north, south, east, west. For the navigation task, the map boundaries are considered as obstacles, while for the manipulation task the cells on the left and right boundaries and top and bottom boundaries are connected to each other since angles are circular. The ground truth shortest path distances are calculated using the Dijkstra algorithm (Dijkstra et al., 1959). Unreachable locations and obstacles are denoted by $- 1$ in the ground truth.
|
| 92 |
+
|
| 93 |
+
In addition to testing on unseen maps with the same distribution, we also test the spatial planning models on two types of out-of-distribution datasets: More Obstacles where we generate $o _ { m i n } = 1 5$ to $o _ { m a x } = 2 0$ obstacles per map, and Real-world where the top-down maps are generated from reconstructions of real-world scenes in Gibson dataset (Xia et al., 2018).
|
| 94 |
+
|
| 95 |
+
Hyperparameters and Training. For training the SPT model, we use Stochastic Gradient Descent (Bottou, 2010) for optimization with a starting learning rate of 1.0 and a learning rate decay of 0.9 per epoch. We train the model for 40 epochs with a batch size of 20. We use $N = 5$ Transformer layers each with $h = 8$ attention heads and a embedding size of $d = 6 4$ . The inner dimension of the fully connected layers in the transformer is $d _ { f c } = 5 1 2$ . We use the same architecture with the same hyperparameters for training the SPT model for both navigation and manipulation for all map sizes. We will open-source the code including dataset generation, model implementation and training.
|
| 96 |
+
|
| 97 |
+
<table><tr><td></td><td colspan="3">Navigation</td><td colspan="2">Manipulation</td><td rowspan="2">Overall</td></tr><tr><td>Method</td><td>M=15</td><td>M=30</td><td>M=50</td><td>M=18</td><td>M=36</td></tr><tr><td>VIN</td><td>86.19</td><td>83.62</td><td>80.84</td><td>75.06</td><td>74.27</td><td>80.00</td></tr><tr><td>GPPN</td><td>97.10</td><td>96.17</td><td>91.97</td><td>89.06</td><td>87.23</td><td>92.31</td></tr><tr><td>SPT</td><td>99.07</td><td>99.56</td><td>99.42</td><td>99.24</td><td>99.78</td><td>99.41</td></tr></table>
|
| 98 |
+
|
| 99 |
+
Table 1: Generalization to in-distribution maps. Table showing the average planning accuracy of the proposed model Spatial Planning Transformer (SPT) as compared to the baselines on in-distribution test sets for both the navigation and manipulation experiments.
|
| 100 |
+
|
| 101 |
+
<table><tr><td></td><td colspan="6">Navigation</td><td colspan="2">Manipulation</td><td>Overall</td></tr><tr><td></td><td colspan="3">More Obstacles</td><td colspan="3">Real-World</td><td colspan="2">More Obstacles</td><td></td></tr><tr><td>Method</td><td>M=15</td><td>M=30</td><td>M=50</td><td>M=15</td><td>M=30</td><td>M=50</td><td>M=18</td><td>M=36</td><td></td></tr><tr><td>VIN</td><td>49.05</td><td>62.05</td><td>70.64</td><td>49.91</td><td>56.67</td><td>71.16</td><td>65.27</td><td>59.81</td><td>60.57</td></tr><tr><td>GPPN</td><td>90.68</td><td>89.93</td><td>84.86</td><td>90.11</td><td>91.07</td><td>88.32</td><td>79.86</td><td>80.79</td><td>86.95</td></tr><tr><td>SPT</td><td>93.34</td><td>92.71</td><td>92.03</td><td>95.96</td><td>94.70</td><td>95.39</td><td>98.16</td><td>99.18</td><td>95.18</td></tr></table>
|
| 102 |
+
|
| 103 |
+
Table 2: Generalization to out-of-distribution maps. Table showing the average planning accuracy of the proposed model Spatial Planning Transformer (SPT) as compared to the baselines on out-of-distribution test sets for both the navigation and manipulation experiments.
|
| 104 |
+
|
| 105 |
+
Baselines. We use prior spatial planning models as baselines. These include Value Iteration Networks (VIN) (Tamar et al., 2016) and Gated Path-Planning Networks (GPPN) (Lee et al., 2018). For tuning the hyperparameter $( K )$ for the number of iterations in both the baselines, we consider all values of $K$ in multiples of 10 such that the inference time of the baseline is comparable to the inference time of the SPT model $\leq 1 . 1$ times). For each setting, we tune $K$ and the learning rate to maximize performance on the validation set.
|
| 106 |
+
|
| 107 |
+
Metrics. We use average action prediction accuracy as the metric. Distance prediction is converted to actions by finding the minimum distance cell among the 4 neighboring cells for each location. When multiple actions are optimal, predicting any optimal action is considered to be a correct prediction. The accuracy is averaged over all free space locations over all maps in the test set.
|
| 108 |
+
|
| 109 |
+
Results. The planning accuracy of all the methods for both the navigation and manipulation tasks on the in-distribution test sets are shown in Table 1 and on the out-of-distribution test sets are shown in Table 2. The proposed SPT model outperforms both the baselines across all settings achieving an overall accuracy of $9 9 . 4 1 \%$ vs $9 2 . 3 1 \%$ (in-distribution) and $9 5 . 1 8 \%$ vs $8 6 . 9 5 \%$ (out-of-distribution) as compared to the best baseline. The performance of the SPT model is stable as the map size increases while the performance of the baselines drop considerably. We believe this is because both the baselines need to use a larger number of iterations to cover a larger map ( $K = 6 0$ iterations for GPPN and $K = 9 0$ iterations for VIN for $M = 5 0$ ) since the information propagation is local in VIN and GPPN. The optimization becomes difficult for such deep models. In contrast, the SPT model uses a constant $N = 5$ layers for all map sizes.
|
| 110 |
+
|
| 111 |
+
The improvement in the performance of SPT over the baselines is larger in the manipulation task because the baselines based on convolution operations are not well suited for propagating information looping over the edges of the map. In contrast, the SPT model can use self-attention to attend to any part of the map and learn to propagate information over the map edges.
|
| 112 |
+
|
| 113 |
+
Visualizations. In Figure 5, we show examples of predictions of the SPT model as compared to the baselines for 3 different input maps and goals from 3 different test sets. The examples show that the baselines are not able to predict the distances of distant cells accurately. This is because they propagate information in a local neighborhood that can not reach distant cells in the limited inference time budget $K = 3 0$ for VIN and $K = 2 0$ for GPPN). In contrast, the SPT is able to predict distances of distant cells more accurately with $N = 5$ layers indicating that it learns long-range information propagation. Additional examples are provided in the Appendix.
|
| 114 |
+
|
| 115 |
+
<table><tr><td colspan="3">Runtime j permapi in ms</td></tr><tr><td>Method</td><td>M=15 M=30</td><td>M=50</td></tr><tr><td>Dijkstra</td><td>4.17</td><td>43.82 371.05</td></tr><tr><td>A</td><td>3.02 35.38</td><td>294.70</td></tr><tr><td>SPT</td><td>2.44 4.72</td><td>18.35</td></tr></table>
|
| 116 |
+
|
| 117 |
+
Table 3: Runtime comparison. Comparison of average runtime per map in milli seconds for different methods. All values are averaged over 10000 maps.
|
| 118 |
+
|
| 119 |
+
Runtime Comparison. To demonstrate one of the benefits of learning-based planners over classical planning algorithms, we compare the runtime of SPT to Dijkstra (Dijkstra et al., 1959) and $\mathbf { A } ^ { * }$ (Hart et al., 1968) algorithms in Table 3. The results indicate that SPT is $1 . 2 4 \times$ to $2 0 . 2 2 \times$ faster than classical planning algorithms with the runtime benefit improving with the increase in map size.
|
| 120 |
+
|
| 121 |
+

|
| 122 |
+
Figure 5: Spatial Planning Examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for map size $M = 3 0$ . The obstacles are shown in blue, free space in purple and goal in yellow in the leftmost input column. The predictions and ground truth in the rest of the column are color-coded from blue to yellow to represent increasing action distance.
|
| 123 |
+
|
| 124 |
+
# 5 EXPERIMENTS & RESULTS: END-TO-END MAPPING AND PLANNING
|
| 125 |
+
|
| 126 |
+
In the above experiments, we compared the planning performance of different methods under perfect knowledge of the map $m$ . In this section, we test the efficacy of spatial planning methods when map $m$ is unknown and needs to be predicted from sensor observations $o$ .
|
| 127 |
+
|
| 128 |
+
Datasets. For manipulation, we generate synthetic datasets of size $M = 1 8$ using the same process as described in Section 4. We discretize the operation space into a $P \times P$ image with $P = 9 0$ which is used as the observation $o$ . The train/test sets are of size 100K/5K.
|
| 129 |
+
|
| 130 |
+
For navigation, we use the Gibson dataset (Xia et al., 2018) to sample maps of size $M = 1 5$ where each cell is $0 . 2 5 m ^ { 2 }$ area. We get the camera images at the navigable locations in all 4 orientations using the Habitat simulator (Savva et al., 2019). The set of camera images each of size $3 \times H \times W$ act as the observation $o$ for the navigation task, where $H = W = 1 2 8$ . The train and test sets consist of 72 and 14 distinct scenes identical to the standard train and val splits in the Habitat simulator. We sample 500 maps in each scene creating training/test sets of size 36K/7K. Each sampled map is rotated to a random orientation.
|
| 131 |
+
|
| 132 |
+
Training. We load the weights of different models trained on synthetic data from the previous section. We then train the end-to-end model using the same action distance prediction loss while keeping the planner weights frozen. The architecture of the mapper module is identical across different planning methods. Metrics. We report both map accuracy and planning accuracy for both the tasks.
|
| 133 |
+
|
| 134 |
+
Baselines. In addition to using VIN and GPPN as baselines, we also use a classical mapping and planning baseline for the navigation task. Since there is no depth input available, we used Monocular depth estimation model from Hu et al. (2019) for predicting the map which is then used for planning using Dijkstra as suggested by Mishkin et al. (2019).
|
| 135 |
+
|
| 136 |
+
Results. Table 4 shows the end-to-end mapping and planning results. SPT outperforms both GPPN and VIN by a large margin across both the tasks achieving an overall plan accuracy of $8 2 . 2 9 \%$ vs $6 3 . 9 1 \%$ . Table 4 also shows that the mapper learnt using end-to-end training with a pretrained SPT model is able to achieve an accuracy of $9 8 . 9 6 \%$ for manipulation and $8 2 . 5 8 \%$ for navigation, without receiving any map-level supervision. SPT also outperforms the classical mapping and planning baseline. These results demonstrate a key benefit of learning-based differentiable planners as compared to classical analytical planning algorithms. As SPT outperforms VIN and GPPN at spatial planning, it also leads to a better map accuracy $( 9 0 . 7 7 \%$ vs $7 7 . 3 1 \%$ ).
|
| 137 |
+
|
| 138 |
+
<table><tr><td></td><td colspan="2">Navigation</td><td colspan="2">Manipulation</td><td colspan="2">Overall</td></tr><tr><td>Method</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td></tr><tr><td>Classical</td><td>64.43</td><td>45.20</td><td></td><td>=</td><td>1</td><td>=</td></tr><tr><td>VIN</td><td>60.92</td><td>47.77</td><td>81.25</td><td>66.45</td><td>71.08</td><td>57.11</td></tr><tr><td>GPPN</td><td>69.06</td><td>45.70</td><td>85.57</td><td>82.13</td><td>77.31</td><td>63.91</td></tr><tr><td>SPT</td><td>82.58</td><td>66.16</td><td>98.96</td><td>98.42</td><td>90.77</td><td>82.29</td></tr></table>
|
| 139 |
+
|
| 140 |
+
Table 4: End-to-End Mapping and Planning Results. Table showing the average mapping and planning accuracy of the proposed model Spatial Planning Transformer (SPT) as compared to the baselines for end-to-end mapping and planning experiments.
|
| 141 |
+
|
| 142 |
+
# 5.1 SPARSE AND NOISY SUPERVISION
|
| 143 |
+
|
| 144 |
+
In the above experiments, we assumed access to perfect and dense action-distance supervision. In practice, if we were to get supervision from human trajectories, the supersion could be sparse, as we might not have access to the optimal distance from all locations in the map, and noisy as humans might not take the optimal actions always and computing distances from human trajectories might be noisy. To study the effect of not having dense and perfect supervision for training the end-to-end mapping and planning model, we consider three settings:
|
| 145 |
+
|
| 146 |
+
Noisy supervision: We add zero mean gaussian noise to all ground-truth distance values with standard deviation, $\sigma = 1$ .
|
| 147 |
+
|
| 148 |
+
Sparse supervision: Instead of providing distances from all navigable locations in the ground truth, we provide distances for only 5 trajectories to the same goal in the training maps.
|
| 149 |
+
|
| 150 |
+
oisy and Sparse supervision: We provide noisy distances for only 5 trajectories as supervision
|
| 151 |
+
|
| 152 |
+
Figure 6 shows an example of noisy and sparse supervision. The results are shown in Table 5. The SPT model maintains performance benefits over the baselines under all the settings. Interestingly, under sparse supervision, the map prediction accuracy drops, but the planning accuracy does not drop as much. This is because the model learns to predict minimum map required to predict the action distances of all valid locations accurately as seen in examples shown in Figure 15 in the Appendix.
|
| 153 |
+
|
| 154 |
+
# 6 RELATED WORK
|
| 155 |
+
|
| 156 |
+
Path planning in known or inferred maps, also known as motion planning in robotics, is a well explored problem led by the seminal papers (Canny, 1988; Kavraki et al., 1996; LaValle & Kuffner Jr, 2001; Karaman & Frazzoli, 2011). Although there are learned variants of motion planners proposed in the literature using gaussian processes (Ijspeert et al., 2013; Ratliff et al., 2018), data-driven motion planners using neural networks is a recent direction (Qureshi et al., 2019; Bhardwaj et al., 2020; Qureshi et al., 2020). Prior work has also studied the use of neural networks to learn the heuristics and sampling stratergies in classical planners Ichter et al. (2018); Guez et al. (2018); Satorras & Welling (2020); Khan et al. (2020). Learning for planning is more common in Markov Decision Process (MDPs) for computing value function via dynamic programming based value iterations (Bellman, 1966; Bertsekas et al., 1995). Planning and learning in neural networks has been explored (Ilin et al., 2007) with a successful general formulation provided by value iteration networks (VIN) (Tamar et al., 2016) with follow-ups to improve scalability and efficiency (Lee et al., 2018; Karkus et al., 2017; Nardelli et al., 2018; Schleich et al., 2019; Khan et al., 2018; Chen et al., 2020). However, these models only capture local value propagation using CNNs and are mostly applied in navigation setups. In contrast, proposed SPTs capture long-range spatial dependency and easily scale to both navigation and manipulation.
|
| 157 |
+
|
| 158 |
+
Differentiable planning structure has also been explored in reinforcement learning with model-free methods (Silver et al., 2017; Oh et al., 2017; Zhu et al., 2017; Farquhar et al., 2018) as well as off-policy RL (Eysenbach et al., 2019; Laskin et al., 2020). Recent works also backpropagate through learned planners to train the policy (Pathak et al., 2018; Srinivas et al., 2018; Amos et al., 2018) and use imagined rollouts of a learned world model for long-term plans (Racaniere et al. \` , 2017; Hafner et al., 2019; Sekar et al., 2019). Unlike our work, these works lack the structure of a spatial planner.
|
| 159 |
+
|
| 160 |
+

|
| 161 |
+
Figure 6: Sparse and Noisy Supervision. Figure showing examples of a map and goal with different levels of supervision. Noisy supervision adds gaussian noise to the ground truth distance values, and sparse supervision samples 5 trajectories for random starting locations to the goal location.
|
| 162 |
+
|
| 163 |
+
<table><tr><td></td><td colspan="2">Dense and Perfect supervision</td><td colspan="2">Noisy supervision</td><td colspan="2">Sparse supervision</td><td colspan="2">Noisy and Sparse supervision</td></tr><tr><td>Method</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td></tr><tr><td>VIN</td><td>81.25</td><td>66.45</td><td>75.68</td><td>60.78</td><td>70.16</td><td>60.23</td><td>70.22</td><td>58.97</td></tr><tr><td>GPPN</td><td>85.57</td><td>82.13</td><td>80.13</td><td>76.11</td><td>72.73</td><td>75.13</td><td>70.08</td><td>72.85</td></tr><tr><td>SPT</td><td>98.96</td><td>98.42</td><td>96.35</td><td>95.83</td><td>80.15</td><td>97.18</td><td>77.17</td><td>94.34</td></tr></table>
|
| 164 |
+
|
| 165 |
+
Table 5: Sparse and Noisy Supervision Results. Table showing the average mapping and planning accuracy of the proposed model Spatial Planning Transformer (SPT) as compared to the baselines for end-to-end mapping and planning experiments under noisy and sparse supervision settings for the manipulation task.
|
| 166 |
+
|
| 167 |
+
Decomposing learning a controller into mapping and planning is common in robot navigation (Khatib, 1986; Elfes, 1987). Some works have explored joint mapping and planning (Elfes, 1989; Fraundorfer et al., 2012). Maps can also be built from vision (Konolige et al., 2010; Fuentes-Pacheco et al., 2015) with a learned mapper (Parisotto & Salakhutdinov, 2017; Karkus et al., 2020). There has been some work on learning maps without using map annotations as well (Gregor et al., 2019). For navigation specific applications, recent works proposed joint mapping and planning for navigation (Gupta et al., 2017; Zhang et al., 2017; Savinov et al., 2018; Chaplot et al., 2020). However, most of these works either require access to ground truth map or assume interaction. Hence, they will first need to be trained in simulation. In contrast, we show results when the map is not known to the agent by learning just from trajectories and can be directly learned from data collected in the real-world.
|
| 168 |
+
|
| 169 |
+
# 7 DISCUSSION
|
| 170 |
+
|
| 171 |
+
The SPT model is designed to learn long-range spatial planning and it outperforms baselines across multiple experimental settings on both navigation and manipulation tasks. End-to-end learning experiments demonstrate that the SPT model can deal with unknown maps by learning mapping without any map-supervision. However, there are some limitations that need to be addressed before learning-based planning methods can be used for large-scale applications. We showed that the SPT model scales much better with increasing map sizes as compared to the baselines, however larger map sizes lead to a high inference time. In the future, we plan to tackle larger map sizes by learning an encoding which reduces the size of the map before spatial planning. The action space was limited to 4 actions in our experiments. Actions can be more fine-grained with smaller cells in the map, which can also be incorporated by tackling larger map sizes in the future.
|
| 172 |
+
|
| 173 |
+
# REFERENCES
|
| 174 |
+
|
| 175 |
+
Brandon Amos, Ivan Jimenez, Jacob Sacks, Byron Boots, and J Zico Kolter. Differentiable mpc for end-to-end planning and control. In NeurIPS, 2018.
|
| 176 |
+
|
| 177 |
+
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.
|
| 178 |
+
|
| 179 |
+
Richard Bellman. Dynamic programming. Science, 1966.
|
| 180 |
+
|
| 181 |
+
Dimitri P Bertsekas, Dimitri P Bertsekas, Dimitri P Bertsekas, and Dimitri P Bertsekas. Dynamic programming and optimal control. Athena scientific Belmont, MA, 1995.
|
| 182 |
+
|
| 183 |
+
Mohak Bhardwaj, Byron Boots, and Mustafa Mukadam. Differentiable gaussian process motion planning. In International Conference on Robotics and Automation (ICRA), 2020.
|
| 184 |
+
|
| 185 |
+
Leon Bottou. Large-scale machine learning with stochastic gradient descent. In ´ Proceedings of COMPSTAT’2010. Springer, 2010.
|
| 186 |
+
|
| 187 |
+
John Canny. The complexity of robot motion planning. MIT press, 1988.
|
| 188 |
+
|
| 189 |
+
Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. arXiv preprint arXiv:2004.05155, 2020.
|
| 190 |
+
|
| 191 |
+
Binghong Chen, Bo Dai, Qinjie Lin, Guo Ye, Han Liu, and Le Song. Learning to plan in high dimensions via neural exploration-exploitation trees. In ICLR, 2020.
|
| 192 |
+
|
| 193 |
+
Edsger W Dijkstra et al. A note on two problems in connexion with graphs. Numerische mathematik, 1959.
|
| 194 |
+
|
| 195 |
+
Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 2006.
|
| 196 |
+
|
| 197 |
+
Alberto Elfes. Sonar-based real-world mapping and navigation. IEEE Journal on Robotics and Automation, 1987.
|
| 198 |
+
|
| 199 |
+
Alberto Elfes. Using occupancy grids for mobile robot perception and navigation. Computer, 1989.
|
| 200 |
+
|
| 201 |
+
Ben Eysenbach, Russ R Salakhutdinov, and Sergey Levine. Search on the replay buffer: Bridging planning and reinforcement learning. In NeurIPS, 2019.
|
| 202 |
+
|
| 203 |
+
Gregory Farquhar, Tim Rocktaschel, Maximilian Igl, and Shimon Whiteson. Treeqn and atreec: ¨ Differentiable tree-structured models for deep reinforcement learning. In ICLR, 2018.
|
| 204 |
+
|
| 205 |
+
Friedrich Fraundorfer, Lionel Heng, Dominik Honegger, Gim Hee Lee, Lorenz Meier, Petri Tanskanen, and Marc Pollefeys. Vision-based autonomous mapping and exploration using a quadrotor mav. In International Conference on Intelligent Robots and Systems, 2012.
|
| 206 |
+
|
| 207 |
+
Jorge Fuentes-Pacheco, Jose Ruiz-Ascencio, and Juan Manuel Rend ´ on-Mancha. Visual simultaneous ´ localization and mapping: a survey. Artificial intelligence review, 2015.
|
| 208 |
+
|
| 209 |
+
Tobias Glasmachers. Limits of end-to-end learning. arXiv preprint arXiv:1704.08305, 2017.
|
| 210 |
+
|
| 211 |
+
Reginald G Golledge, Valerie Dougherty, and Scott Bell. Acquiring spatial knowledge: Survey versus route-based knowledge in unfamiliar environments. Annals of the association of American geographers, 1995.
|
| 212 |
+
|
| 213 |
+
Karol Gregor, Danilo Jimenez Rezende, Frederic Besse, Yan Wu, Hamza Merzic, and Aaron van den Oord. Shaping belief states with generative environment models for rl. Advances in Neural Information Processing Systems, 32:13475–13487, 2019.
|
| 214 |
+
|
| 215 |
+
Arthur Guez, Theophane Weber, Ioannis Antonoglou, Karen Simonyan, Oriol Vinyals, Daan Wierstra, Remi Munos, and David Silver. Learning to search with mctsnets. In International Conference on Machine Learning, pp. 1822–1831, 2018.
|
| 216 |
+
|
| 217 |
+
Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.
|
| 218 |
+
|
| 219 |
+
Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In ICML, 2019.
|
| 220 |
+
|
| 221 |
+
Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100–107, 1968.
|
| 222 |
+
|
| 223 |
+
Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1043–1051. IEEE, 2019.
|
| 224 |
+
|
| 225 |
+
Brian Ichter, James Harrison, and Marco Pavone. Learning sampling distributions for robot motion planning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 7087–7094. IEEE, 2018.
|
| 226 |
+
|
| 227 |
+
Auke Jan Ijspeert, Jun Nakanishi, Heiko Hoffmann, Peter Pastor, and Stefan Schaal. Dynamical movement primitives: learning attractor models for motor behaviors. Neural computation, 2013.
|
| 228 |
+
|
| 229 |
+
Roman Ilin, Robert Kozma, and Paul J Werbos. Efficient learning in cellular simultaneous recurrent neural networks-the case of maze navigation problem. In 2007 IEEE International Symposium on Approximate Dynamic Programming and Reinforcement Learning, 2007.
|
| 230 |
+
|
| 231 |
+
Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 2011.
|
| 232 |
+
|
| 233 |
+
Peter Karkus, David Hsu, and Wee Sun Lee. Qmdp-net: Deep learning for planning under partial observability. In Advances in Neural Information Processing Systems, 2017.
|
| 234 |
+
|
| 235 |
+
Peter Karkus, Anelia Angelova, Vincent Vanhoucke, and Rico Jonschkowski. Differentiable mapping networks: Learning structured map representations for sparse visual localization. arXiv preprint arXiv:2005.09530, 2020.
|
| 236 |
+
|
| 237 |
+
Lydia E Kavraki, Petr Svestka, J-C Latombe, and Mark H Overmars. Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE transactions on Robotics and Automation, 1996.
|
| 238 |
+
|
| 239 |
+
Arbaaz Khan, Clark Zhang, Nikolay Atanasov, Konstantinos Karydis, Vijay Kumar, and Daniel D Lee. Memory augmented control networks. In ICLR, 2018.
|
| 240 |
+
|
| 241 |
+
Arbaaz Khan, Alejandro Ribeiro, Vijay Kumar, and Anthony G Francis. Graph neural networks for motion planning. arXiv preprint arXiv:2006.06248, 2020.
|
| 242 |
+
|
| 243 |
+
Oussama Khatib. Real-time obstacle avoidance for manipulators and mobile robots. In Autonomous robot vehicles. Springer, 1986.
|
| 244 |
+
|
| 245 |
+
Kurt Konolige, James Bowman, JD Chen, Patrick Mihelich, Michael Calonder, Vincent Lepetit, and Pascal Fua. View-based maps. The International Journal of Robotics Research, 2010.
|
| 246 |
+
|
| 247 |
+
Michael Laskin, Scott Emmons, Ajay Jain, Thanard Kurutach, Pieter Abbeel, and Deepak Pathak. Sparse graphical memory for robust planning. arXiv preprint arXiv:2003.06417, 2020.
|
| 248 |
+
|
| 249 |
+
Steven M LaValle and James J Kuffner Jr. Randomized kinodynamic planning. The international journal of robotics research, 2001.
|
| 250 |
+
|
| 251 |
+
Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 1998.
|
| 252 |
+
|
| 253 |
+
Lisa Lee, Emilio Parisotto, Devendra Singh Chaplot, Eric Xing, and Ruslan Salakhutdinov. Gated path planning networks. In ICML, 2018.
|
| 254 |
+
|
| 255 |
+
Tomas Lozano-Perez. Spatial planning: A configuration space approach. In Autonomous robot vehicles. Springer, 1990.
|
| 256 |
+
|
| 257 |
+
Dmytro Mishkin, Alexey Dosovitskiy, and Vladlen Koltun. Benchmarking classic and learned navigation in complex 3d environments. arXiv preprint arXiv:1901.10915, 2019.
|
| 258 |
+
|
| 259 |
+
Nantas Nardelli, Gabriel Synnaeve, Zeming Lin, Pushmeet Kohli, Philip HS Torr, and Nicolas Usunier. Value propagation networks. arXiv preprint arXiv:1805.11199, 2018.
|
| 260 |
+
|
| 261 |
+
Junhyuk Oh, Satinder Singh, and Honglak Lee. Value prediction network. In Advances in Neural Information Processing Systems, 2017.
|
| 262 |
+
|
| 263 |
+
Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. arXiv preprint arXiv:1702.08360, 2017.
|
| 264 |
+
|
| 265 |
+
Deepak Pathak, Parsa Mahmoudieh, Guanghao Luo, Pulkit Agrawal, Dian Chen, Yide Shentu, Evan Shelhamer, Jitendra Malik, Alexei A Efros, and Trevor Darrell. Zero-shot visual imitation. In ICLR, 2018.
|
| 266 |
+
|
| 267 |
+
Ahmed H Qureshi, Anthony Simeonov, Mayur J Bency, and Michael C Yip. Motion planning networks. In International Conference on Robotics and Automation (ICRA), 2019.
|
| 268 |
+
|
| 269 |
+
Ahmed H Qureshi, Jiangeng Dong, Austin Choe, and Michael C Yip. Neural manipulation planning on constraint manifolds. IEEE Robotics and Automation Letters, 5(4):6089–6096, 2020.
|
| 270 |
+
|
| 271 |
+
Sebastien Racani ´ ere, Th \` eophane Weber, David Reichert, Lars Buesing, Arthur Guez, Danilo Jimenez ´ Rezende, Adria Puigdomenech Badia, Oriol Vinyals, Nicolas Heess, Yujia Li, et al. Imaginationaugmented agents for deep reinforcement learning. In NIPS, 2017.
|
| 272 |
+
|
| 273 |
+
Nathan D Ratliff, Jan Issac, Daniel Kappler, Stan Birchfield, and Dieter Fox. Riemannian motion policies. arXiv preprint arXiv:1801.02854, 2018.
|
| 274 |
+
|
| 275 |
+
Victor Garcia Satorras and Max Welling. Neural enhanced belief propagation on factor graphs. arXiv preprint arXiv:2003.01998, 2020.
|
| 276 |
+
|
| 277 |
+
Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. In ICLR, 2018.
|
| 278 |
+
|
| 279 |
+
Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019.
|
| 280 |
+
|
| 281 |
+
Daniel Schleich, Tobias Klamt, and Sven Behnke. Value iteration networks on multiple levels of abstraction. arXiv preprint arXiv:1905.11068, 2019.
|
| 282 |
+
|
| 283 |
+
Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In ICML, 2019.
|
| 284 |
+
|
| 285 |
+
David Silver, Hado Hasselt, Matteo Hessel, Tom Schaul, Arthur Guez, Tim Harley, Gabriel DulacArnold, David Reichert, Neil Rabinowitz, Andre Barreto, et al. The predictron: End-to-end learning and planning. In International Conference on Machine Learning. PMLR, 2017.
|
| 286 |
+
|
| 287 |
+
Aravind Srinivas, Allan Jabri, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Universal planning networks. ICML, 2018.
|
| 288 |
+
|
| 289 |
+
Aviv Tamar, Yi Wu, Garrett Thomas, Sergey Levine, and Pieter Abbeel. Value iteration networks. In NIPS, 2016.
|
| 290 |
+
|
| 291 |
+
Edward C Tolman. Cognitive maps in rats and men. Psychological review, 1948.
|
| 292 |
+
|
| 293 |
+
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017.
|
| 294 |
+
|
| 295 |
+
Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In CVPR, 2018.
|
| 296 |
+
|
| 297 |
+
Jingwei Zhang, Lei Tai, Joschka Boedecker, Wolfram Burgard, and Ming Liu. Neural slam: Learning to explore with external memory. arXiv preprint arXiv:1706.09520, 2017.
|
| 298 |
+
|
| 299 |
+
Yuke Zhu, Daniel Gordon, Eric Kolve, Dieter Fox, Li Fei-Fei, Abhinav Gupta, Roozbeh Mottaghi, and Ali Farhadi. Visual semantic planning using deep successor representations. In ICCV, 2017.
|
| 300 |
+
|
| 301 |
+
# A BACKGROUND: TRANSFORMERS
|
| 302 |
+
|
| 303 |
+
The proposed spatial planning method is based on the Transformer model (Vaswani et al., 2017). A Transformer layer, denoted by $f _ { \mathrm { T L } }$ , takes a tensor $X \in \mathcal { R } ^ { d \times S }$ as input, where $d$ is the embedding size and $S$ is the size of the input. It consists of two sublayers, a multi-head self-attention layer $( f _ { \mathrm { S A } } )$ and a position-wise fully connected layer $( f _ { \mathrm { F C } } )$ . There is a residual connection around each sublayer, followed by layer normalization (Ba et al., 2016) (LN):
|
| 304 |
+
|
| 305 |
+
$$
|
| 306 |
+
R = \operatorname { L N } ( f _ { \operatorname { S A } } ( X ) + X ) , \quad Y = f _ { \operatorname { T L } } ( X ) = \operatorname { L N } ( f _ { \operatorname { F C } } ( R ) + R )
|
| 307 |
+
$$
|
| 308 |
+
|
| 309 |
+
where $R , Y \in \mathcal { R } ^ { d \times S }$ are the intermediate and final representations, respectively.
|
| 310 |
+
|
| 311 |
+
The multi-head self-attention $( f _ { \mathrm { S A } } )$ layer has $h$ attention heads, each computes a scaled dot-product attention over queries $Q$ , keys $K$ and values $V$ , which are all different projections of the input $X$ :
|
| 312 |
+
|
| 313 |
+
$$
|
| 314 |
+
Q _ { i } = W _ { Q , i } ^ { T } X , \quad K _ { i } = W _ { K , i } ^ { T } X , \quad V = W _ { V , i } ^ { T } X
|
| 315 |
+
$$
|
| 316 |
+
|
| 317 |
+
$$
|
| 318 |
+
Z _ { i } = \mathrm { A t t e n t i o n } ( Q _ { i } , K _ { i } , V _ { i } ) = \mathrm { s o f t m a x } \left( \frac { Q _ { i } K _ { i } ^ { T } } { \sqrt { d _ { k } } } \right) V _ { i }
|
| 319 |
+
$$
|
| 320 |
+
|
| 321 |
+
where $Q , K \in \mathcal { R } ^ { d _ { k } \times S }$ , $V \in \mathcal { R } ^ { d _ { v } \times S }$ , $i \in { 1 , 2 , \dots , h \ d _ { k } }$ and $d _ { v }$ are hyper-parameters and all $W \mathrm { { s } }$ are parameters. The output of all attention heads, $Z _ { i } \mathrm { s }$ , are concatenated and projected to the same dimension as the input. Finally, the position-wise fully connected $( f _ { \mathrm { F C } } )$ layer applies two linear transformations to each position with a ReLU activations to the output of the multi-head attention.
|
| 322 |
+
|
| 323 |
+
# B DATASET DETAILS
|
| 324 |
+
|
| 325 |
+
We generate synthetic datasets for training the spatial planning models for both navigation and manipulation settings. For the navigation setting, we perform experiments with $M \times M$ maps with two different map sizes, $M \in \{ 1 5 , 3 0 \}$ . We randomly generate $o _ { m i n } = 0$ to $o _ { m a x } = 5$ obstacles in each map, where each obstacle is an rectangle at a random location with each side being a random length from 1 to $M / 2$ . All the rectangular obstacles are rotated in two random orientations.
|
| 326 |
+
|
| 327 |
+
For the manipulation setting, we consider a reacher task using a planar arm with 2 degrees of freedom. We use an operational space of size $P \times P$ . Each link of the arm is of size $P / 4$ . The arm is centered at the center of the operational space. Let the orientation of two links be denoted by $\theta _ { 1 }$ and $\theta _ { 2 }$ . We assume both the links can freely rotate in a plane, $\theta _ { 1 } , \theta _ { 2 } \in [ 0 , 2 \pi )$ . For each environment, we generate $o _ { m i n } = 0$ to $o _ { m a x } = 5$ circular obstacles centered at a random location $0 . 2 5 P$ to $0 . 7 5 P$ distance away from the center, with a random radius between $0 . 0 5 P$ and $D - 0 . 1 5 P$ where $D$ is the distance of the center of the obstacle from the center of the operational space. We convert each environment to a configuration space map of size $M \times M$ , where each cell $( i , j )$ denotes whether the arm will collide with an obstacle when $\theta _ { 1 } = 2 \pi i / M$ and $\theta _ { 2 } = 2 \pi j / M$ . We experiment with two map sizes, $M \in \{ 1 8 , 3 6 \}$ , corresponding to $2 0 ^ { \circ }$ and $1 0 ^ { \circ }$ bins for each link. The choice of $P$ does not affect the map as the collision check for each cell in the configuration space is performed in the continuous operational space where all distances are relative to $P$ .
|
| 328 |
+
|
| 329 |
+
# C NAVIGATION MAPPER ARCHITECTURE DETAILS
|
| 330 |
+
|
| 331 |
+
The Navigation Mapper module predicts a single value between 0 and 1 for each image in $o$ indicating whether the cell in the front of the image is an obstacle or not. The architecture of the Navigation mapper consists of ResNet18 convolutional layers followed by 3 fully-connected layers of size 256, 128 and 1 as shown in Figure 7. Each cell can have upto 4 predictions (from images corresponding to the four neighboring cells facing the current cell), which are aggregated using max pooling to get a single prediction.
|
| 332 |
+
|
| 333 |
+
# D HIGHER DIMENSIONAL STATE AND ACTION SPACES
|
| 334 |
+
|
| 335 |
+
To test whether SPT maintains performance benefits in higher dimensional state and action spaces, we conducted some experiments for the navigation task. We relaxed the action space from 4 actions to 100 actions by just allowing the agent to take any action in a $1 0 \mathrm { x } 1 0$ grid around it (and using a low-level controller to go to any cell). The state space used for planning is discretized but the agent moves in a continuous state space in the Habitat simulator. We compute the continuous ground truth distance using Fast Marching Method (instead of Dijkstra) for training with a larger action space, which allows us to accurately compute distance for all locations and not be constrained by axis-aligned actions and distances.
|
| 336 |
+
|
| 337 |
+

|
| 338 |
+
Figure 7: Navigation Mapper Architecture. Figure showing the architecture of the Navigation Mapper.
|
| 339 |
+
|
| 340 |
+
SPT and the baselines are trained only on synthetic navigation mazes for this experiment. During evaluation, we assume perfect map based on part of the environment seen in the observations so far for planning. If the overall map size at this level of discretization is higher than planning map size, we simply use greedy planning in a window around the agent resulting in an “anytime” variant just like the classical RRT algorithms. This setup results in much finer-grained action space as shown in the demo example here 1. SPT achieve navigation success rate of $7 8 . 0 \%$ as compared to $4 7 . 2 \%$ for GPPN and $4 3 . 5 \%$ for VIN baselines.
|
| 341 |
+
|
| 342 |
+
# E ATTENTION VISUALIZATION
|
| 343 |
+
|
| 344 |
+
We show the visualization of attention map corresponding to two different locations in Figure 8. Interestingly, we noticed three consistent patterns: a) at least one of the attention head out of eight captures obstacles (left), b) one of the attention heads focuses on goal location (middle), and c) some attention maps focus on nearby obstacles to get accurate planning distance (right).
|
| 345 |
+
|
| 346 |
+

|
| 347 |
+
Figure 8: Attention Visualization. Visualization of the attention heads learned by Spatial Planning Transformers. SPTs learn an attention for each location in the map with respect to every other location.
|
| 348 |
+
|
| 349 |
+
# F EXAMPLES
|
| 350 |
+
|
| 351 |
+
We show additional examples for navigation task for in-distribution test set (in Figure 9), out-ofdistribution More Obstacles test set (in Figure 10) and Real-World test set (in Figure 11) each with map size $M = 3 0$ . Additional examples for manipulation task are shown for in-distribution test set (in Figure 12) and for out-of-distribution More Obstacles test set (in Figure 13).
|
| 352 |
+
|
| 353 |
+
We also visualize examples for the end-to-end mapping and planning experiments for the manipulation task. We show examples of map and action distance predictions using the SPT model trained with dense and perfect supervision in Figure 14 and with noisy and sparse supervision in Figure 15.
|
| 354 |
+
|
| 355 |
+

|
| 356 |
+
Figure 9: Navigation in-distribution test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Navigation in-distribution test set for map size $M = 3 0$ .
|
| 357 |
+
|
| 358 |
+

|
| 359 |
+
Figure 10: Navigation out-of-distribution More Obstacles test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Navigation out-of-distribution More Obstacles test set for map size $M = 3 0$ .
|
| 360 |
+
|
| 361 |
+

|
| 362 |
+
Figure 11: Navigation out-of-distribution Real-World test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Navigation out-of-distribution Real-World test set for map size $M = 3 0$ .
|
| 363 |
+
|
| 364 |
+

|
| 365 |
+
Figure 12: Manipulation in-distribution test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Manipulation indistribution test set for map size $M = 3 6$ .
|
| 366 |
+
|
| 367 |
+

|
| 368 |
+
Figure 13: Manipulation out-of-distribution More Obstacles test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Manipulation out-of-distribution More Obstacles test set for map size $M = 3 6$ .
|
| 369 |
+
|
| 370 |
+

|
| 371 |
+
Figure 14: Dense and Perfect Supervision. Figure showing examples of map and distance predictions using the SPT model trained with dense and perfect action-level supervision.
|
| 372 |
+
|
| 373 |
+

|
| 374 |
+
Figure 15: Sparse and Noisy Supervision. Figure showing examples of map and distance predictions using the SPT model trained with sparse and noisy action-level supervision.
|
parse/train/n4IMHNb8_f/n4IMHNb8_f_content_list.json
ADDED
|
@@ -0,0 +1,2032 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "DIFFERENTIABLE SPATIAL PLANNING USING TRANSFORMERS ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
174,
|
| 8 |
+
99,
|
| 9 |
+
629,
|
| 10 |
+
146
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Anonymous authors Paper under double-blind review ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
183,
|
| 19 |
+
171,
|
| 20 |
+
398,
|
| 21 |
+
198
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "ABSTRACT ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
454,
|
| 31 |
+
222,
|
| 32 |
+
544,
|
| 33 |
+
238
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "We consider the problem of spatial path planning. In contrast to the classical solutions which optimize a new plan from scratch and assume access to the full map with ground truth obstacle locations, we learn a planner from the data in a differentiable manner that allows us to leverage statistical regularities from past data. We propose Spatial Planning Transformers (SPT), which given an obstacle map learns to generate actions by planning over long-range spatial dependencies, unlike prior data-driven planners that propagate information locally via convolutional structure in an iterative manner. In the setting where the ground truth map is not known to the agent, we leverage pre-trained SPTs to in an end-to-end framework that has the structure of mapper and planner built into it which allows seamless generalization to out-of-distribution maps and goals. SPTs outperform prior stateof-the-art across all the setups for both manipulation and navigation tasks, leading to an absolute improvement of $7 . 1 9 \\%$ . ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
233,
|
| 42 |
+
260,
|
| 43 |
+
766,
|
| 44 |
+
439
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "1 INTRODUCTION ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
176,
|
| 54 |
+
473,
|
| 55 |
+
336,
|
| 56 |
+
489
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "The problem of path planning has been a bedrock of robotics. Given an obstacle map of an environment and a goal location in the map, the task is to output a shortest path to the goal location starting from any position in the map. We consider path planning with spatial maps. Building a top-down spatial map is common practice in robotic navigation as it provides a natural representation of physical space (Durrant-Whyte & Bailey, 2006). In fact, even Observations robotic manipulation can also be naturally phrased via spatial map using the formalism of configuration spaces (Lozano-Perez, 1990), as shown in Figure 1. This problem has been studied in robotics for several decades, and classic goto planning algorithms involve Dijkstra et al. (1959), PRM (Kavraki et al., 1996), RRT (LaValle & Kuffner Jr, 2001), RRT\\* (Karaman & Frazzoli, 2011), etc. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
501,
|
| 66 |
+
509,
|
| 67 |
+
679
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "image",
|
| 73 |
+
"img_path": "images/7f5a10b005711d04ad7a1735e55486427173da448d048d7b2898969de738601b.jpg",
|
| 74 |
+
"image_caption": [
|
| 75 |
+
"Figure 1: Spatial Path Planning: The raw obserPerception Planning Map Actionsvations (top left) and obstacles can be represented spatially via top-down map in navigation (left) and 3via configuration space in manipulation (right). "
|
| 76 |
+
],
|
| 77 |
+
"image_footnote": [],
|
| 78 |
+
"bbox": [
|
| 79 |
+
524,
|
| 80 |
+
502,
|
| 81 |
+
823,
|
| 82 |
+
607
|
| 83 |
+
],
|
| 84 |
+
"page_idx": 0
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"type": "text",
|
| 88 |
+
"text": "",
|
| 89 |
+
"bbox": [
|
| 90 |
+
176,
|
| 91 |
+
681,
|
| 92 |
+
823,
|
| 93 |
+
708
|
| 94 |
+
],
|
| 95 |
+
"page_idx": 0
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"type": "text",
|
| 99 |
+
"text": "Our objective is to develop methods that can learn to plan from data. However, a natural question is why do we need learning for a problem which has stable classical solutions? There are two key reasons. First, classical methods do not capture statistical regularities present in the natural world, (for e.g., walls are mostly parallel or perpendicular to each other), because they optimize a plan from scratch for each new setup. This also makes analytical planning methods to be often slow at inference time which is an issue in dynamic scenarios where a more reactive policy might be required for fast adaptation from failures. A learned planner represented via a neural network can not only capture regularities but is also efficient at inference as the plan is just a result of forward-pass through the network. Second, a critical assumption of classical algorithms is that a global ground-truth obstacle space must be known to the agent ahead of time. This is in stark contrast to biological agents where cognitive maps are not pixel-accurate ground truth location of agents, but built through actions in the environment, e.g., rats build an implicit map of the environment incrementally through trajectories enabling them to take shortcuts (Tolman, 1948). A learned solution could not only provides the ability to deal with partial, noisy maps and but also help build maps on the fly while acting in the environment by backpropagating through the generated long-range plans. ",
|
| 100 |
+
"bbox": [
|
| 101 |
+
174,
|
| 102 |
+
715,
|
| 103 |
+
825,
|
| 104 |
+
924
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 0
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "image",
|
| 110 |
+
"img_path": "images/d3a44ec22e9fafe4a214e9a9f98206e0ae81c3ea50180cbc3b1d651c3eb8d97b.jpg",
|
| 111 |
+
"image_caption": [
|
| 112 |
+
"Figure 2: Local vs Long-distance value propagation. Figure showing an example of number of iterations required to propagate distance values over a map using local and long-distance value propagation. The obstacle map and goal location shown on the left and the distance value predictions over 5 iterations is shown on the right (distance values increase from blue to yellow). Prior methods based on convolutional networks use local value propagation and require many iterations to propagate values accurately over the whole map (top right). Our method is based on long-distance value propagation between points without any obstacle between them. This type of value propagation can cover the whole map in 3 iterations in this example (bottom right). "
|
| 113 |
+
],
|
| 114 |
+
"image_footnote": [],
|
| 115 |
+
"bbox": [
|
| 116 |
+
199,
|
| 117 |
+
98,
|
| 118 |
+
820,
|
| 119 |
+
279
|
| 120 |
+
],
|
| 121 |
+
"page_idx": 1
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"type": "text",
|
| 125 |
+
"text": "Several recent works have proposed data-driven path planning models (Tamar et al., 2016; Karkus et al., 2017; Nardelli et al., 2018; Lee et al., 2018). Similar to how classical algorithms, like Dijkstra et al. (1959), move outward from the goal one cell at a time to predict distances iteratively based on the obstacles in the map, current learning-based spatial planning models propagate distance values in only a local neighborhood using convolutional networks. This kind of local value propagation requires $\\mathcal { O } ( M )$ iterations, where $M$ is the map dimension. In theory, however, the optimal paths can be computed much more efficiently with total iterations that are on the order of number of obstacles rather than the map size. For instance, consider two points with no obstacle between, an efficient planner could directly connect them with interpolated distance. Nonetheless, this is possible only if the model can perform long-range reasoning in the obstacle space which is a challenge. ",
|
| 126 |
+
"bbox": [
|
| 127 |
+
173,
|
| 128 |
+
392,
|
| 129 |
+
825,
|
| 130 |
+
531
|
| 131 |
+
],
|
| 132 |
+
"page_idx": 1
|
| 133 |
+
},
|
| 134 |
+
{
|
| 135 |
+
"type": "text",
|
| 136 |
+
"text": "In this work, our goal is to capture this long-range spatial relationship. Transformers (Vaswani et al., 2017) are well suited for this kind of computation as they treat the inputs as sets and propagate information across all the points within the set. Building on this, we propose Spatial Planning Transformers (SPT) which consists of attention heads that can attend to any part of the input. The key idea behind the design of the proposed model is that value can be propagated between distant points if there are no obstacles between them. This would reduce the number of required iterations to $\\mathcal { O } ( n _ { O } )$ where $n _ { O }$ is the number of obstacles in the map. Figure 2 shows a simple example where long-distance value propagation can cover the entire map within 3 iterations while local value propagation takes more than 5 iterations – this difference grows with the complexity of the obstacle space and map size. We compare the performance of SPTs with prior state-of-the-art learned planning approaches, VIN (Tamar et al., 2016) and GPPN (Lee et al., 2018), across both navigation as well as manipulation setups. SPTs achieve significantly higher accuracy than these prior methods for the same inference time and show over $1 0 \\%$ absolute improvement when the maps are large. ",
|
| 137 |
+
"bbox": [
|
| 138 |
+
173,
|
| 139 |
+
537,
|
| 140 |
+
825,
|
| 141 |
+
718
|
| 142 |
+
],
|
| 143 |
+
"page_idx": 1
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"type": "text",
|
| 147 |
+
"text": "Next, we turn to the case when the map is not known apriori. This is a practical setting when the agent either has access to a partially known map or just know it through the trajectories. In psychology, this is known as going from route knowledge to survey knowledege (Golledge et al., 1995) where animals aggregate the knowledge from trajectories into a cognitive map. We operationalize this setup by formulating an end-to-end differentiable framework, which in contrast to having a generic parametric policy learning (Glasmachers, 2017), has the structure of mapper and planner built into it. We first pre-train the SPT planner to capture a generic data-driven prior, and then backpropagate through it to learn a mapper that maps raw observations to obstacle map. This allows us to learn without requiring map supervision or interaction. Learned mapper and planner not only allow us to plan for new goal locations at inference but also generalize to unseen maps. ",
|
| 148 |
+
"bbox": [
|
| 149 |
+
174,
|
| 150 |
+
724,
|
| 151 |
+
825,
|
| 152 |
+
864
|
| 153 |
+
],
|
| 154 |
+
"page_idx": 1
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"type": "text",
|
| 158 |
+
"text": "Our end-to-end mapping and planning approach provides a unified solution for both navigation and manipulation. We perform thorough experiments in both manipulation and real-world navigation maps as well as manipulation. Our approach outperforms prior state-of-the-art by a margin on both mapping and planning accuracy without assuming access to the map at training or inference. ",
|
| 159 |
+
"bbox": [
|
| 160 |
+
174,
|
| 161 |
+
872,
|
| 162 |
+
823,
|
| 163 |
+
928
|
| 164 |
+
],
|
| 165 |
+
"page_idx": 1
|
| 166 |
+
},
|
| 167 |
+
{
|
| 168 |
+
"type": "image",
|
| 169 |
+
"img_path": "images/7fee31e5309b671e59d1065612e02cc88f5cdf9ec9d07aea3416189280e3b93d.jpg",
|
| 170 |
+
"image_caption": [
|
| 171 |
+
"Figure 3: Spatial Planning Transformer (SPT). Figure showing an overview of the proposed Spatial Planning Transformer model. It consists of 3 modules: an Encoder $E$ to encode the input, a Transformer network $T$ responsible for planning, and a Decoder $D$ decoding the output of the Transformer into action distances. "
|
| 172 |
+
],
|
| 173 |
+
"image_footnote": [],
|
| 174 |
+
"bbox": [
|
| 175 |
+
176,
|
| 176 |
+
97,
|
| 177 |
+
823,
|
| 178 |
+
352
|
| 179 |
+
],
|
| 180 |
+
"page_idx": 2
|
| 181 |
+
},
|
| 182 |
+
{
|
| 183 |
+
"type": "text",
|
| 184 |
+
"text": "2 PRELIMINARIES AND PROBLEM DEFINITION ",
|
| 185 |
+
"text_level": 1,
|
| 186 |
+
"bbox": [
|
| 187 |
+
174,
|
| 188 |
+
411,
|
| 189 |
+
573,
|
| 190 |
+
428
|
| 191 |
+
],
|
| 192 |
+
"page_idx": 2
|
| 193 |
+
},
|
| 194 |
+
{
|
| 195 |
+
"type": "text",
|
| 196 |
+
"text": "We represent the input spatial map as a matrix, $m$ , of size $M \\times M$ with each element being 1, denoting obstacles, or 0, denoting free space. The goal location is also represented as a matrix, $g$ , of size $M \\times M$ with exactly one element being 1, denoting the goal location, and rest 0s. The input to the spatial planning model, $x$ , consists of matrices $m$ and $g$ stacked, $x = [ m , g ]$ , where $x$ is of size $2 \\times M \\times M$ . The objective of the planning model is to predict $y$ which is of size $M \\times M$ , consisting of action distances of corresponding locations from the goal. Here, action distance is defined to be the minimum number of actions required to reach the goal. ",
|
| 197 |
+
"bbox": [
|
| 198 |
+
173,
|
| 199 |
+
434,
|
| 200 |
+
825,
|
| 201 |
+
531
|
| 202 |
+
],
|
| 203 |
+
"page_idx": 2
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"type": "text",
|
| 207 |
+
"text": "For navigation, $m$ is a top-down obstacle map, and $g$ represents the goal position on this map. For manipulation, $m$ represents the obstacles in the configuration space of 2-dof planar arm with joint angles denoted by $\\theta _ { 1 }$ and $\\theta _ { 2 }$ . Each element $( i , j )$ in $m$ indicate whether the configuration of the arm with joint angles $\\theta _ { 1 } = i$ and $\\theta _ { 2 } = j$ , would lead to a collision. $g$ represents the goal configuration of the arm. In the first set of experiments, we will assume that $m$ is known and in the second set of experiments, $m$ is not known and the agent receives observations, $o$ , from its sensors instead. ",
|
| 208 |
+
"bbox": [
|
| 209 |
+
174,
|
| 210 |
+
537,
|
| 211 |
+
825,
|
| 212 |
+
622
|
| 213 |
+
],
|
| 214 |
+
"page_idx": 2
|
| 215 |
+
},
|
| 216 |
+
{
|
| 217 |
+
"type": "text",
|
| 218 |
+
"text": "3 METHODS ",
|
| 219 |
+
"text_level": 1,
|
| 220 |
+
"bbox": [
|
| 221 |
+
174,
|
| 222 |
+
636,
|
| 223 |
+
292,
|
| 224 |
+
651
|
| 225 |
+
],
|
| 226 |
+
"page_idx": 2
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"type": "text",
|
| 230 |
+
"text": "We design a spatial planning model capable of long-distance information propagation. We first describe the design of this spatial planning module, called Spatial Planning Transformer(SPT), shown in Figure 3, which takes in a map and a goal as input and predict the distance to the goal from all locations. We then describe how the SPT model can be used as a planning module to train end-to-end learning models, which take in raw sensory observations and goal location as input and predict action distances without having access to the map. ",
|
| 231 |
+
"bbox": [
|
| 232 |
+
174,
|
| 233 |
+
657,
|
| 234 |
+
825,
|
| 235 |
+
741
|
| 236 |
+
],
|
| 237 |
+
"page_idx": 2
|
| 238 |
+
},
|
| 239 |
+
{
|
| 240 |
+
"type": "text",
|
| 241 |
+
"text": "3.1 SPT: SPATIAL PLANNING TRANSFORMERS ",
|
| 242 |
+
"text_level": 1,
|
| 243 |
+
"bbox": [
|
| 244 |
+
174,
|
| 245 |
+
757,
|
| 246 |
+
511,
|
| 247 |
+
772
|
| 248 |
+
],
|
| 249 |
+
"page_idx": 2
|
| 250 |
+
},
|
| 251 |
+
{
|
| 252 |
+
"type": "text",
|
| 253 |
+
"text": "To propogate information over distant points, we use the Transformer (Vaswani et al., 2017) architecture. The self-attention mechanism in a Transformer can learn to attend to any element of the input. The allows the model to learn spatial reasoning over the whole map concurrently. Figure 3 shows an overview of the SPT model, which consists of three modules, an Encoder $E$ to encode the input, a Transformer network $T$ responsible for spatial planning, and a Decoder $D$ decoding the output of the Transformer into action distances. ",
|
| 254 |
+
"bbox": [
|
| 255 |
+
174,
|
| 256 |
+
777,
|
| 257 |
+
825,
|
| 258 |
+
861
|
| 259 |
+
],
|
| 260 |
+
"page_idx": 2
|
| 261 |
+
},
|
| 262 |
+
{
|
| 263 |
+
"type": "text",
|
| 264 |
+
"text": "Encoder. The Encoder $E$ computes the encoding of the input $x$ : $x _ { I } ~ = ~ E ( x )$ . The input x ∈ {0, 1}2×M×M consisting of the map and goal is first passed through a 2-layer convolutional network (LeCun et al., 1998) with ReLU activations to compute an embedding for each input element. Both layers have a kernel size of $1 \\times 1$ , which ensures that the embedding of all the obstacles is identical to each other, and the same holds true for free space and the goal location. The output of this convolutional network is of size $d \\times M \\times M$ , where $d$ is the embedding size. This output is then flattened to get $x _ { I }$ of size $d \\times M ^ { 2 }$ and passed into the Transformer network. ",
|
| 265 |
+
"bbox": [
|
| 266 |
+
174,
|
| 267 |
+
867,
|
| 268 |
+
825,
|
| 269 |
+
924
|
| 270 |
+
],
|
| 271 |
+
"page_idx": 2
|
| 272 |
+
},
|
| 273 |
+
{
|
| 274 |
+
"type": "image",
|
| 275 |
+
"img_path": "images/c5893cc0a08552573389574a5a0a25e71f67c54b7ff1080e5b3a25a059c28251.jpg",
|
| 276 |
+
"image_caption": [
|
| 277 |
+
"Figure 4: End-to-end Mapping and Planning. Figure showing an overview of end-to-end mapping and planning model for both the navigation and manipulation tasks. "
|
| 278 |
+
],
|
| 279 |
+
"image_footnote": [],
|
| 280 |
+
"bbox": [
|
| 281 |
+
184,
|
| 282 |
+
92,
|
| 283 |
+
813,
|
| 284 |
+
290
|
| 285 |
+
],
|
| 286 |
+
"page_idx": 3
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"type": "text",
|
| 290 |
+
"text": "",
|
| 291 |
+
"bbox": [
|
| 292 |
+
174,
|
| 293 |
+
337,
|
| 294 |
+
825,
|
| 295 |
+
380
|
| 296 |
+
],
|
| 297 |
+
"page_idx": 3
|
| 298 |
+
},
|
| 299 |
+
{
|
| 300 |
+
"type": "text",
|
| 301 |
+
"text": "Transformer. The Transformer network $T$ converts the input encoding into the output encoding: $x _ { O } = T ( x _ { I } )$ . It first adds the positional encoding to the input encoding. The positional encoding enables the Transformer model to distinguish between the obstacles at different locations. We use a constant sinusoidal positional encoding (Vaswani et al., 2017): ",
|
| 302 |
+
"bbox": [
|
| 303 |
+
174,
|
| 304 |
+
386,
|
| 305 |
+
825,
|
| 306 |
+
443
|
| 307 |
+
],
|
| 308 |
+
"page_idx": 3
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"type": "equation",
|
| 312 |
+
"img_path": "images/eda727b07069fa52c2a17047ab4b0dd319df7b48a853e7c16c72c55be211581d.jpg",
|
| 313 |
+
"text": "$$\np _ { ( 2 i , j ) } = \\sin ( j / C ^ { 2 i / d } ) , \\qquad p _ { ( 2 i + 1 , j ) } = \\cos ( j / C ^ { 2 i / d } )\n$$",
|
| 314 |
+
"text_format": "latex",
|
| 315 |
+
"bbox": [
|
| 316 |
+
318,
|
| 317 |
+
446,
|
| 318 |
+
678,
|
| 319 |
+
469
|
| 320 |
+
],
|
| 321 |
+
"page_idx": 3
|
| 322 |
+
},
|
| 323 |
+
{
|
| 324 |
+
"type": "text",
|
| 325 |
+
"text": "where $p \\in \\mathcal { R } ^ { d \\times M ^ { 2 } }$ is the positional encoding, $j \\in \\{ 1 , 2 , \\dots , M ^ { 2 } \\}$ is the position of the input, $i \\in \\{ 1 , 2 , \\ldots , d / 2 \\}$ , and $C \\bar { = } M ^ { 2 }$ is a constant. ",
|
| 326 |
+
"bbox": [
|
| 327 |
+
174,
|
| 328 |
+
483,
|
| 329 |
+
821,
|
| 330 |
+
513
|
| 331 |
+
],
|
| 332 |
+
"page_idx": 3
|
| 333 |
+
},
|
| 334 |
+
{
|
| 335 |
+
"type": "text",
|
| 336 |
+
"text": "The positional encoding of each element is added to their corresponding input encoding to get $Z = x _ { I } + p . Z$ is then passed through $N = 5$ identical Transformer layers $( f _ { \\mathrm { T L } } )$ to get $x _ { O }$ . ",
|
| 337 |
+
"bbox": [
|
| 338 |
+
174,
|
| 339 |
+
520,
|
| 340 |
+
821,
|
| 341 |
+
549
|
| 342 |
+
],
|
| 343 |
+
"page_idx": 3
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"type": "text",
|
| 347 |
+
"text": "Decoder. The Decoder $D$ computes the distance prediction $\\hat { y }$ from $x _ { O }$ using a position-wise fully connected layer: ",
|
| 348 |
+
"bbox": [
|
| 349 |
+
166,
|
| 350 |
+
554,
|
| 351 |
+
821,
|
| 352 |
+
583
|
| 353 |
+
],
|
| 354 |
+
"page_idx": 3
|
| 355 |
+
},
|
| 356 |
+
{
|
| 357 |
+
"type": "equation",
|
| 358 |
+
"img_path": "images/6bbca5979a3365ba77550b2e6bbd86957d89e2a516e2dd6c6544cbde6e3aac50.jpg",
|
| 359 |
+
"text": "$$\n\\hat { y } _ { i } = W _ { D } ^ { T } x _ { T , i } + b _ { D }\n$$",
|
| 360 |
+
"text_format": "latex",
|
| 361 |
+
"bbox": [
|
| 362 |
+
433,
|
| 363 |
+
582,
|
| 364 |
+
565,
|
| 365 |
+
599
|
| 366 |
+
],
|
| 367 |
+
"page_idx": 3
|
| 368 |
+
},
|
| 369 |
+
{
|
| 370 |
+
"type": "text",
|
| 371 |
+
"text": "where $\\boldsymbol { x } _ { T , i } \\in \\mathcal { R } ^ { d \\times 1 }$ is the input at position $i \\in { 1 , 2 , \\dots , M ^ { 2 } }$ , $W _ { D } \\in \\mathcal { R } ^ { d \\times 1 } , b _ { D } \\in \\mathcal { R }$ are parameters of the Decoder shared across all positions $i$ and $\\hat { y } _ { i } \\in \\mathcal { R }$ is the distance prediction at position $i$ . The distance prediction at all position are reshaped into a matrix to get the final prediction $\\hat { y } \\in \\mathcal { R } ^ { M , M }$ . The entire model is trained using pairs of input $x$ and output $y$ datapoints with mean-squared error as the loss function. ",
|
| 372 |
+
"bbox": [
|
| 373 |
+
173,
|
| 374 |
+
603,
|
| 375 |
+
826,
|
| 376 |
+
674
|
| 377 |
+
],
|
| 378 |
+
"page_idx": 3
|
| 379 |
+
},
|
| 380 |
+
{
|
| 381 |
+
"type": "text",
|
| 382 |
+
"text": "3.2 END-TO-END MAPPING AND PLANNING ",
|
| 383 |
+
"text_level": 1,
|
| 384 |
+
"bbox": [
|
| 385 |
+
176,
|
| 386 |
+
690,
|
| 387 |
+
491,
|
| 388 |
+
705
|
| 389 |
+
],
|
| 390 |
+
"page_idx": 3
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"type": "text",
|
| 394 |
+
"text": "The SPT model described above is designed to predict action distances given a map as input. However, in many applications, the map of the environment is often not known. In such cases, an autonomous agent working in a realistic environment needs to predict the map from raw sensory observations. While it is possible to train a separate mapper model to predict maps from observations, this often requires map annotations which are expensive to obtain and often inaccurate. In contrast, demonstration trajectories consisting of observations and optimal actions are more readily available or easier to obtain in many robotics applications. One of the key benefits of learning-based differentiable spatial planning is that it can be used to learn mapping just from action supervision in an end-to-end fashion without having access to ground-truth maps. To demonstrate this benefit, we train an endto-end mapping and planning model to predict action distances from sensor observations for both navigation and manipulation tasks. ",
|
| 395 |
+
"bbox": [
|
| 396 |
+
173,
|
| 397 |
+
709,
|
| 398 |
+
826,
|
| 399 |
+
863
|
| 400 |
+
],
|
| 401 |
+
"page_idx": 3
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"type": "text",
|
| 405 |
+
"text": "The end-to-end mapping and planning model consists of two modules, a Mapper $( f _ { M } )$ and a Planner $( f _ { P } )$ , as illustrated in Figure 4. The Mapper is used to predict the map $\\hat { m }$ from sensor observations $o$ and the Planner is a spatial planning model to predict action distances, $\\hat { y }$ , from the predicted map $\\hat { m }$ : ",
|
| 406 |
+
"bbox": [
|
| 407 |
+
174,
|
| 408 |
+
869,
|
| 409 |
+
825,
|
| 410 |
+
912
|
| 411 |
+
],
|
| 412 |
+
"page_idx": 3
|
| 413 |
+
},
|
| 414 |
+
{
|
| 415 |
+
"type": "equation",
|
| 416 |
+
"img_path": "images/4a266fe79fd70dbc2b390b50176558fb91017b8c661af98438e57deecee6e643.jpg",
|
| 417 |
+
"text": "$$\n\\hat { y } = f _ { P } ( \\hat { m } ) = f _ { P } ( f _ { M } ( o ) )\n$$",
|
| 418 |
+
"text_format": "latex",
|
| 419 |
+
"bbox": [
|
| 420 |
+
410,
|
| 421 |
+
917,
|
| 422 |
+
586,
|
| 423 |
+
934
|
| 424 |
+
],
|
| 425 |
+
"page_idx": 3
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"type": "text",
|
| 429 |
+
"text": "For navigation, $o$ is the set of first-person RGB camera images each of size $3 \\times H \\times W$ . We sample 4 images, one for each orientation, at each valid location in the map. For invalid locations, we pass an empty image of 0s for all orientations. Thus, for a map of size $M \\times M$ , observation $o$ consist of $4 M ^ { 2 }$ images for all locations and 4 orientations similar to the setup in Lee et al. (2018). For manipulation, $o$ is a top-down view of the operational space with obstacles of size $P \\times P$ , where each element is 1 or 0 denoting obstacles or free space. We use different Mapper architectures for navigation and manipulation experiments. ",
|
| 430 |
+
"bbox": [
|
| 431 |
+
173,
|
| 432 |
+
103,
|
| 433 |
+
825,
|
| 434 |
+
202
|
| 435 |
+
],
|
| 436 |
+
"page_idx": 4
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"type": "text",
|
| 440 |
+
"text": "The Navigation Mapper module predicts a single value between 0 and 1 for each image in $o$ indicating whether the cell in the front of the image is an obstacle or not. The architecture of the Navigation mapper consists of ResNet18 convolutional layers followed by fully-connected layers (see Appendix for details). Each cell can have upto 4 predictions (from images corresponding to the four neighboring cells facing the current cell), which are aggregated using max pooling to get a single prediction. This map prediction consisting of continuous values is passed to the Planner module. ",
|
| 441 |
+
"bbox": [
|
| 442 |
+
174,
|
| 443 |
+
208,
|
| 444 |
+
825,
|
| 445 |
+
291
|
| 446 |
+
],
|
| 447 |
+
"page_idx": 4
|
| 448 |
+
},
|
| 449 |
+
{
|
| 450 |
+
"type": "text",
|
| 451 |
+
"text": "The Manipulation Mapper module needs to predict which configurations of the arm would lead to a collision. To predict whether a particular configuration $( \\theta _ { 1 } , \\theta _ { 2 } )$ , the mapper module needs to check whether any point in this configuration consists of an obstacle. A Transformer-based model is well suited to learn this function as well as it can attend to arbitrary locations in the operational space to predict the obstacles in the configuration space. We use the same architecture of the SPT model as the Manipulation Mapper as well, with the only difference being the encoder consisting of $3 \\times 3$ kernel size convolutional layers instead of $1 \\times 1$ to encode the $P \\times P$ observation space to a $M \\times M$ representation. ",
|
| 452 |
+
"bbox": [
|
| 453 |
+
174,
|
| 454 |
+
297,
|
| 455 |
+
825,
|
| 456 |
+
410
|
| 457 |
+
],
|
| 458 |
+
"page_idx": 4
|
| 459 |
+
},
|
| 460 |
+
{
|
| 461 |
+
"type": "text",
|
| 462 |
+
"text": "The Planner module is the entire SPT model with encoder, transformer and decoder units as described in the previous subsection. It is pretrained on synthetic maps and its weights are frozen during end-to-end training. We train the entire end-to-end Mapping and Planning model with pairs of input observations $o$ and output action distances $y$ using standard supervised learning with the meansquared error as the loss function: $\\begin{array} { r } { \\mathcal { L } = M S E ( y , \\hat { y } ) = 1 / | L | \\sum _ { i \\in L } ( y _ { i } - \\hat { y } _ { i } ) ^ { 2 } } \\end{array}$ , where $L$ is the set of navigable locations. Since the planning module is pretrained and it expects a structured map input, the mapper model needs to predict the map accurately such that the predicted map, when passed through the planner, minimizes the action level loss. ",
|
| 463 |
+
"bbox": [
|
| 464 |
+
174,
|
| 465 |
+
416,
|
| 466 |
+
825,
|
| 467 |
+
529
|
| 468 |
+
],
|
| 469 |
+
"page_idx": 4
|
| 470 |
+
},
|
| 471 |
+
{
|
| 472 |
+
"type": "text",
|
| 473 |
+
"text": "4 EXPERIMENTS & RESULTS: SPATIAL PLANNING ",
|
| 474 |
+
"text_level": 1,
|
| 475 |
+
"bbox": [
|
| 476 |
+
174,
|
| 477 |
+
553,
|
| 478 |
+
606,
|
| 479 |
+
569
|
| 480 |
+
],
|
| 481 |
+
"page_idx": 4
|
| 482 |
+
},
|
| 483 |
+
{
|
| 484 |
+
"type": "text",
|
| 485 |
+
"text": "Datasets. We generate synthetic datasets for training the spatial planning models for both navigation and manipulation settings. For the navigation setting, we perform experiments with $M \\times M$ maps with three different map sizes, $M \\in \\{ 1 5 , 3 0 , 5 0 \\}$ . For manipulation, we experiment with two map sizes, $M \\in \\{ 1 8 , 3 6 \\}$ , corresponding to $2 0 ^ { \\circ }$ and $1 0 ^ { \\circ }$ bins for each link. In each map, we randomly generate $o _ { m i n } = 0$ to $o _ { m a x } = 5$ obstacles. Dataset generation details are provided in the Appendix. ",
|
| 486 |
+
"bbox": [
|
| 487 |
+
174,
|
| 488 |
+
582,
|
| 489 |
+
825,
|
| 490 |
+
652
|
| 491 |
+
],
|
| 492 |
+
"page_idx": 4
|
| 493 |
+
},
|
| 494 |
+
{
|
| 495 |
+
"type": "text",
|
| 496 |
+
"text": "For both the settings, we generate training, validation, and test sets of size $1 0 0 K / 5 K / 5 K$ maps. The set of maps in each set are distinct. For each map, we choose a random free space cell as the goal location. The action space consists of 4 actions: north, south, east, west. For the navigation task, the map boundaries are considered as obstacles, while for the manipulation task the cells on the left and right boundaries and top and bottom boundaries are connected to each other since angles are circular. The ground truth shortest path distances are calculated using the Dijkstra algorithm (Dijkstra et al., 1959). Unreachable locations and obstacles are denoted by $- 1$ in the ground truth. ",
|
| 497 |
+
"bbox": [
|
| 498 |
+
174,
|
| 499 |
+
659,
|
| 500 |
+
825,
|
| 501 |
+
756
|
| 502 |
+
],
|
| 503 |
+
"page_idx": 4
|
| 504 |
+
},
|
| 505 |
+
{
|
| 506 |
+
"type": "text",
|
| 507 |
+
"text": "In addition to testing on unseen maps with the same distribution, we also test the spatial planning models on two types of out-of-distribution datasets: More Obstacles where we generate $o _ { m i n } = 1 5$ to $o _ { m a x } = 2 0$ obstacles per map, and Real-world where the top-down maps are generated from reconstructions of real-world scenes in Gibson dataset (Xia et al., 2018). ",
|
| 508 |
+
"bbox": [
|
| 509 |
+
174,
|
| 510 |
+
763,
|
| 511 |
+
825,
|
| 512 |
+
819
|
| 513 |
+
],
|
| 514 |
+
"page_idx": 4
|
| 515 |
+
},
|
| 516 |
+
{
|
| 517 |
+
"type": "text",
|
| 518 |
+
"text": "Hyperparameters and Training. For training the SPT model, we use Stochastic Gradient Descent (Bottou, 2010) for optimization with a starting learning rate of 1.0 and a learning rate decay of 0.9 per epoch. We train the model for 40 epochs with a batch size of 20. We use $N = 5$ Transformer layers each with $h = 8$ attention heads and a embedding size of $d = 6 4$ . The inner dimension of the fully connected layers in the transformer is $d _ { f c } = 5 1 2$ . We use the same architecture with the same hyperparameters for training the SPT model for both navigation and manipulation for all map sizes. We will open-source the code including dataset generation, model implementation and training. ",
|
| 519 |
+
"bbox": [
|
| 520 |
+
174,
|
| 521 |
+
825,
|
| 522 |
+
825,
|
| 523 |
+
924
|
| 524 |
+
],
|
| 525 |
+
"page_idx": 4
|
| 526 |
+
},
|
| 527 |
+
{
|
| 528 |
+
"type": "table",
|
| 529 |
+
"img_path": "images/ff23d6cf223112dffc4316662edb5bb01402d48f0028dfe4faa9fbd02db388f1.jpg",
|
| 530 |
+
"table_caption": [],
|
| 531 |
+
"table_footnote": [
|
| 532 |
+
"Table 1: Generalization to in-distribution maps. Table showing the average planning accuracy of the proposed model Spatial Planning Transformer (SPT) as compared to the baselines on in-distribution test sets for both the navigation and manipulation experiments. "
|
| 533 |
+
],
|
| 534 |
+
"table_body": "<table><tr><td></td><td colspan=\"3\">Navigation</td><td colspan=\"2\">Manipulation</td><td rowspan=\"2\">Overall</td></tr><tr><td>Method</td><td>M=15</td><td>M=30</td><td>M=50</td><td>M=18</td><td>M=36</td></tr><tr><td>VIN</td><td>86.19</td><td>83.62</td><td>80.84</td><td>75.06</td><td>74.27</td><td>80.00</td></tr><tr><td>GPPN</td><td>97.10</td><td>96.17</td><td>91.97</td><td>89.06</td><td>87.23</td><td>92.31</td></tr><tr><td>SPT</td><td>99.07</td><td>99.56</td><td>99.42</td><td>99.24</td><td>99.78</td><td>99.41</td></tr></table>",
|
| 535 |
+
"bbox": [
|
| 536 |
+
174,
|
| 537 |
+
90,
|
| 538 |
+
825,
|
| 539 |
+
176
|
| 540 |
+
],
|
| 541 |
+
"page_idx": 5
|
| 542 |
+
},
|
| 543 |
+
{
|
| 544 |
+
"type": "table",
|
| 545 |
+
"img_path": "images/fb001a369741d2e857ee717bb875fa73b266c6be06054e9a3b9e596aef11098e.jpg",
|
| 546 |
+
"table_caption": [],
|
| 547 |
+
"table_footnote": [
|
| 548 |
+
"Table 2: Generalization to out-of-distribution maps. Table showing the average planning accuracy of the proposed model Spatial Planning Transformer (SPT) as compared to the baselines on out-of-distribution test sets for both the navigation and manipulation experiments. "
|
| 549 |
+
],
|
| 550 |
+
"table_body": "<table><tr><td></td><td colspan=\"6\">Navigation</td><td colspan=\"2\">Manipulation</td><td>Overall</td></tr><tr><td></td><td colspan=\"3\">More Obstacles</td><td colspan=\"3\">Real-World</td><td colspan=\"2\">More Obstacles</td><td></td></tr><tr><td>Method</td><td>M=15</td><td>M=30</td><td>M=50</td><td>M=15</td><td>M=30</td><td>M=50</td><td>M=18</td><td>M=36</td><td></td></tr><tr><td>VIN</td><td>49.05</td><td>62.05</td><td>70.64</td><td>49.91</td><td>56.67</td><td>71.16</td><td>65.27</td><td>59.81</td><td>60.57</td></tr><tr><td>GPPN</td><td>90.68</td><td>89.93</td><td>84.86</td><td>90.11</td><td>91.07</td><td>88.32</td><td>79.86</td><td>80.79</td><td>86.95</td></tr><tr><td>SPT</td><td>93.34</td><td>92.71</td><td>92.03</td><td>95.96</td><td>94.70</td><td>95.39</td><td>98.16</td><td>99.18</td><td>95.18</td></tr></table>",
|
| 551 |
+
"bbox": [
|
| 552 |
+
174,
|
| 553 |
+
223,
|
| 554 |
+
821,
|
| 555 |
+
321
|
| 556 |
+
],
|
| 557 |
+
"page_idx": 5
|
| 558 |
+
},
|
| 559 |
+
{
|
| 560 |
+
"type": "text",
|
| 561 |
+
"text": "Baselines. We use prior spatial planning models as baselines. These include Value Iteration Networks (VIN) (Tamar et al., 2016) and Gated Path-Planning Networks (GPPN) (Lee et al., 2018). For tuning the hyperparameter $( K )$ for the number of iterations in both the baselines, we consider all values of $K$ in multiples of 10 such that the inference time of the baseline is comparable to the inference time of the SPT model $\\leq 1 . 1$ times). For each setting, we tune $K$ and the learning rate to maximize performance on the validation set. ",
|
| 562 |
+
"bbox": [
|
| 563 |
+
173,
|
| 564 |
+
372,
|
| 565 |
+
825,
|
| 566 |
+
457
|
| 567 |
+
],
|
| 568 |
+
"page_idx": 5
|
| 569 |
+
},
|
| 570 |
+
{
|
| 571 |
+
"type": "text",
|
| 572 |
+
"text": "Metrics. We use average action prediction accuracy as the metric. Distance prediction is converted to actions by finding the minimum distance cell among the 4 neighboring cells for each location. When multiple actions are optimal, predicting any optimal action is considered to be a correct prediction. The accuracy is averaged over all free space locations over all maps in the test set. ",
|
| 573 |
+
"bbox": [
|
| 574 |
+
174,
|
| 575 |
+
464,
|
| 576 |
+
825,
|
| 577 |
+
520
|
| 578 |
+
],
|
| 579 |
+
"page_idx": 5
|
| 580 |
+
},
|
| 581 |
+
{
|
| 582 |
+
"type": "text",
|
| 583 |
+
"text": "Results. The planning accuracy of all the methods for both the navigation and manipulation tasks on the in-distribution test sets are shown in Table 1 and on the out-of-distribution test sets are shown in Table 2. The proposed SPT model outperforms both the baselines across all settings achieving an overall accuracy of $9 9 . 4 1 \\%$ vs $9 2 . 3 1 \\%$ (in-distribution) and $9 5 . 1 8 \\%$ vs $8 6 . 9 5 \\%$ (out-of-distribution) as compared to the best baseline. The performance of the SPT model is stable as the map size increases while the performance of the baselines drop considerably. We believe this is because both the baselines need to use a larger number of iterations to cover a larger map ( $K = 6 0$ iterations for GPPN and $K = 9 0$ iterations for VIN for $M = 5 0$ ) since the information propagation is local in VIN and GPPN. The optimization becomes difficult for such deep models. In contrast, the SPT model uses a constant $N = 5$ layers for all map sizes. ",
|
| 584 |
+
"bbox": [
|
| 585 |
+
173,
|
| 586 |
+
526,
|
| 587 |
+
825,
|
| 588 |
+
666
|
| 589 |
+
],
|
| 590 |
+
"page_idx": 5
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"type": "text",
|
| 594 |
+
"text": "The improvement in the performance of SPT over the baselines is larger in the manipulation task because the baselines based on convolution operations are not well suited for propagating information looping over the edges of the map. In contrast, the SPT model can use self-attention to attend to any part of the map and learn to propagate information over the map edges. ",
|
| 595 |
+
"bbox": [
|
| 596 |
+
173,
|
| 597 |
+
672,
|
| 598 |
+
823,
|
| 599 |
+
729
|
| 600 |
+
],
|
| 601 |
+
"page_idx": 5
|
| 602 |
+
},
|
| 603 |
+
{
|
| 604 |
+
"type": "text",
|
| 605 |
+
"text": "Visualizations. In Figure 5, we show examples of predictions of the SPT model as compared to the baselines for 3 different input maps and goals from 3 different test sets. The examples show that the baselines are not able to predict the distances of distant cells accurately. This is because they propagate information in a local neighborhood that can not reach distant cells in the limited inference time budget $K = 3 0$ for VIN and $K = 2 0$ for GPPN). In contrast, the SPT is able to predict distances of distant cells more accurately with $N = 5$ layers indicating that it learns long-range information propagation. Additional examples are provided in the Appendix. ",
|
| 606 |
+
"bbox": [
|
| 607 |
+
174,
|
| 608 |
+
736,
|
| 609 |
+
568,
|
| 610 |
+
888
|
| 611 |
+
],
|
| 612 |
+
"page_idx": 5
|
| 613 |
+
},
|
| 614 |
+
{
|
| 615 |
+
"type": "table",
|
| 616 |
+
"img_path": "images/9014bb683d297b9f9502c5a34ea899488f763b2025eca70215ea6e5ef141213d.jpg",
|
| 617 |
+
"table_caption": [],
|
| 618 |
+
"table_footnote": [
|
| 619 |
+
"Table 3: Runtime comparison. Comparison of average runtime per map in milli seconds for different methods. All values are averaged over 10000 maps. "
|
| 620 |
+
],
|
| 621 |
+
"table_body": "<table><tr><td colspan=\"3\">Runtime j permapi in ms</td></tr><tr><td>Method</td><td>M=15 M=30</td><td>M=50</td></tr><tr><td>Dijkstra</td><td>4.17</td><td>43.82 371.05</td></tr><tr><td>A</td><td>3.02 35.38</td><td>294.70</td></tr><tr><td>SPT</td><td>2.44 4.72</td><td>18.35</td></tr></table>",
|
| 622 |
+
"bbox": [
|
| 623 |
+
583,
|
| 624 |
+
739,
|
| 625 |
+
820,
|
| 626 |
+
832
|
| 627 |
+
],
|
| 628 |
+
"page_idx": 5
|
| 629 |
+
},
|
| 630 |
+
{
|
| 631 |
+
"type": "text",
|
| 632 |
+
"text": "Runtime Comparison. To demonstrate one of the benefits of learning-based planners over classical planning algorithms, we compare the runtime of SPT to Dijkstra (Dijkstra et al., 1959) and $\\mathbf { A } ^ { * }$ (Hart et al., 1968) algorithms in Table 3. The results indicate that SPT is $1 . 2 4 \\times$ to $2 0 . 2 2 \\times$ faster than classical planning algorithms with the runtime benefit improving with the increase in map size. ",
|
| 633 |
+
"bbox": [
|
| 634 |
+
173,
|
| 635 |
+
895,
|
| 636 |
+
825,
|
| 637 |
+
924
|
| 638 |
+
],
|
| 639 |
+
"page_idx": 5
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"type": "image",
|
| 643 |
+
"img_path": "images/739eb457a76bbcf42d9773eed6af9036fc2a61e883a847504fc1dec1a0d71ab2.jpg",
|
| 644 |
+
"image_caption": [
|
| 645 |
+
"Figure 5: Spatial Planning Examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for map size $M = 3 0$ . The obstacles are shown in blue, free space in purple and goal in yellow in the leftmost input column. The predictions and ground truth in the rest of the column are color-coded from blue to yellow to represent increasing action distance. "
|
| 646 |
+
],
|
| 647 |
+
"image_footnote": [],
|
| 648 |
+
"bbox": [
|
| 649 |
+
178,
|
| 650 |
+
94,
|
| 651 |
+
810,
|
| 652 |
+
378
|
| 653 |
+
],
|
| 654 |
+
"page_idx": 6
|
| 655 |
+
},
|
| 656 |
+
{
|
| 657 |
+
"type": "text",
|
| 658 |
+
"text": "",
|
| 659 |
+
"bbox": [
|
| 660 |
+
174,
|
| 661 |
+
448,
|
| 662 |
+
823,
|
| 663 |
+
476
|
| 664 |
+
],
|
| 665 |
+
"page_idx": 6
|
| 666 |
+
},
|
| 667 |
+
{
|
| 668 |
+
"type": "text",
|
| 669 |
+
"text": "5 EXPERIMENTS & RESULTS: END-TO-END MAPPING AND PLANNING ",
|
| 670 |
+
"text_level": 1,
|
| 671 |
+
"bbox": [
|
| 672 |
+
171,
|
| 673 |
+
488,
|
| 674 |
+
774,
|
| 675 |
+
506
|
| 676 |
+
],
|
| 677 |
+
"page_idx": 6
|
| 678 |
+
},
|
| 679 |
+
{
|
| 680 |
+
"type": "text",
|
| 681 |
+
"text": "In the above experiments, we compared the planning performance of different methods under perfect knowledge of the map $m$ . In this section, we test the efficacy of spatial planning methods when map $m$ is unknown and needs to be predicted from sensor observations $o$ . ",
|
| 682 |
+
"bbox": [
|
| 683 |
+
174,
|
| 684 |
+
512,
|
| 685 |
+
825,
|
| 686 |
+
554
|
| 687 |
+
],
|
| 688 |
+
"page_idx": 6
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"type": "text",
|
| 692 |
+
"text": "Datasets. For manipulation, we generate synthetic datasets of size $M = 1 8$ using the same process as described in Section 4. We discretize the operation space into a $P \\times P$ image with $P = 9 0$ which is used as the observation $o$ . The train/test sets are of size 100K/5K. ",
|
| 693 |
+
"bbox": [
|
| 694 |
+
174,
|
| 695 |
+
560,
|
| 696 |
+
825,
|
| 697 |
+
602
|
| 698 |
+
],
|
| 699 |
+
"page_idx": 6
|
| 700 |
+
},
|
| 701 |
+
{
|
| 702 |
+
"type": "text",
|
| 703 |
+
"text": "For navigation, we use the Gibson dataset (Xia et al., 2018) to sample maps of size $M = 1 5$ where each cell is $0 . 2 5 m ^ { 2 }$ area. We get the camera images at the navigable locations in all 4 orientations using the Habitat simulator (Savva et al., 2019). The set of camera images each of size $3 \\times H \\times W$ act as the observation $o$ for the navigation task, where $H = W = 1 2 8$ . The train and test sets consist of 72 and 14 distinct scenes identical to the standard train and val splits in the Habitat simulator. We sample 500 maps in each scene creating training/test sets of size 36K/7K. Each sampled map is rotated to a random orientation. ",
|
| 704 |
+
"bbox": [
|
| 705 |
+
174,
|
| 706 |
+
609,
|
| 707 |
+
825,
|
| 708 |
+
707
|
| 709 |
+
],
|
| 710 |
+
"page_idx": 6
|
| 711 |
+
},
|
| 712 |
+
{
|
| 713 |
+
"type": "text",
|
| 714 |
+
"text": "Training. We load the weights of different models trained on synthetic data from the previous section. We then train the end-to-end model using the same action distance prediction loss while keeping the planner weights frozen. The architecture of the mapper module is identical across different planning methods. Metrics. We report both map accuracy and planning accuracy for both the tasks. ",
|
| 715 |
+
"bbox": [
|
| 716 |
+
174,
|
| 717 |
+
714,
|
| 718 |
+
825,
|
| 719 |
+
770
|
| 720 |
+
],
|
| 721 |
+
"page_idx": 6
|
| 722 |
+
},
|
| 723 |
+
{
|
| 724 |
+
"type": "text",
|
| 725 |
+
"text": "Baselines. In addition to using VIN and GPPN as baselines, we also use a classical mapping and planning baseline for the navigation task. Since there is no depth input available, we used Monocular depth estimation model from Hu et al. (2019) for predicting the map which is then used for planning using Dijkstra as suggested by Mishkin et al. (2019). ",
|
| 726 |
+
"bbox": [
|
| 727 |
+
174,
|
| 728 |
+
776,
|
| 729 |
+
823,
|
| 730 |
+
833
|
| 731 |
+
],
|
| 732 |
+
"page_idx": 6
|
| 733 |
+
},
|
| 734 |
+
{
|
| 735 |
+
"type": "text",
|
| 736 |
+
"text": "Results. Table 4 shows the end-to-end mapping and planning results. SPT outperforms both GPPN and VIN by a large margin across both the tasks achieving an overall plan accuracy of $8 2 . 2 9 \\%$ vs $6 3 . 9 1 \\%$ . Table 4 also shows that the mapper learnt using end-to-end training with a pretrained SPT model is able to achieve an accuracy of $9 8 . 9 6 \\%$ for manipulation and $8 2 . 5 8 \\%$ for navigation, without receiving any map-level supervision. SPT also outperforms the classical mapping and planning baseline. These results demonstrate a key benefit of learning-based differentiable planners as compared to classical analytical planning algorithms. As SPT outperforms VIN and GPPN at spatial planning, it also leads to a better map accuracy $( 9 0 . 7 7 \\%$ vs $7 7 . 3 1 \\%$ ). ",
|
| 737 |
+
"bbox": [
|
| 738 |
+
174,
|
| 739 |
+
840,
|
| 740 |
+
825,
|
| 741 |
+
924
|
| 742 |
+
],
|
| 743 |
+
"page_idx": 6
|
| 744 |
+
},
|
| 745 |
+
{
|
| 746 |
+
"type": "table",
|
| 747 |
+
"img_path": "images/a992364235e133f26b5051ca9e530b6d9637a24fafeca92c09ebb8947550815f.jpg",
|
| 748 |
+
"table_caption": [],
|
| 749 |
+
"table_footnote": [
|
| 750 |
+
"Table 4: End-to-End Mapping and Planning Results. Table showing the average mapping and planning accuracy of the proposed model Spatial Planning Transformer (SPT) as compared to the baselines for end-to-end mapping and planning experiments. "
|
| 751 |
+
],
|
| 752 |
+
"table_body": "<table><tr><td></td><td colspan=\"2\">Navigation</td><td colspan=\"2\">Manipulation</td><td colspan=\"2\">Overall</td></tr><tr><td>Method</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td></tr><tr><td>Classical</td><td>64.43</td><td>45.20</td><td></td><td>=</td><td>1</td><td>=</td></tr><tr><td>VIN</td><td>60.92</td><td>47.77</td><td>81.25</td><td>66.45</td><td>71.08</td><td>57.11</td></tr><tr><td>GPPN</td><td>69.06</td><td>45.70</td><td>85.57</td><td>82.13</td><td>77.31</td><td>63.91</td></tr><tr><td>SPT</td><td>82.58</td><td>66.16</td><td>98.96</td><td>98.42</td><td>90.77</td><td>82.29</td></tr></table>",
|
| 753 |
+
"bbox": [
|
| 754 |
+
173,
|
| 755 |
+
101,
|
| 756 |
+
826,
|
| 757 |
+
200
|
| 758 |
+
],
|
| 759 |
+
"page_idx": 7
|
| 760 |
+
},
|
| 761 |
+
{
|
| 762 |
+
"type": "text",
|
| 763 |
+
"text": "",
|
| 764 |
+
"bbox": [
|
| 765 |
+
176,
|
| 766 |
+
270,
|
| 767 |
+
823,
|
| 768 |
+
299
|
| 769 |
+
],
|
| 770 |
+
"page_idx": 7
|
| 771 |
+
},
|
| 772 |
+
{
|
| 773 |
+
"type": "text",
|
| 774 |
+
"text": "5.1 SPARSE AND NOISY SUPERVISION ",
|
| 775 |
+
"text_level": 1,
|
| 776 |
+
"bbox": [
|
| 777 |
+
176,
|
| 778 |
+
319,
|
| 779 |
+
449,
|
| 780 |
+
333
|
| 781 |
+
],
|
| 782 |
+
"page_idx": 7
|
| 783 |
+
},
|
| 784 |
+
{
|
| 785 |
+
"type": "text",
|
| 786 |
+
"text": "In the above experiments, we assumed access to perfect and dense action-distance supervision. In practice, if we were to get supervision from human trajectories, the supersion could be sparse, as we might not have access to the optimal distance from all locations in the map, and noisy as humans might not take the optimal actions always and computing distances from human trajectories might be noisy. To study the effect of not having dense and perfect supervision for training the end-to-end mapping and planning model, we consider three settings: ",
|
| 787 |
+
"bbox": [
|
| 788 |
+
174,
|
| 789 |
+
339,
|
| 790 |
+
825,
|
| 791 |
+
422
|
| 792 |
+
],
|
| 793 |
+
"page_idx": 7
|
| 794 |
+
},
|
| 795 |
+
{
|
| 796 |
+
"type": "text",
|
| 797 |
+
"text": "Noisy supervision: We add zero mean gaussian noise to all ground-truth distance values with standard deviation, $\\sigma = 1$ . ",
|
| 798 |
+
"bbox": [
|
| 799 |
+
176,
|
| 800 |
+
424,
|
| 801 |
+
823,
|
| 802 |
+
449
|
| 803 |
+
],
|
| 804 |
+
"page_idx": 7
|
| 805 |
+
},
|
| 806 |
+
{
|
| 807 |
+
"type": "text",
|
| 808 |
+
"text": "Sparse supervision: Instead of providing distances from all navigable locations in the ground truth, we provide distances for only 5 trajectories to the same goal in the training maps. ",
|
| 809 |
+
"bbox": [
|
| 810 |
+
173,
|
| 811 |
+
450,
|
| 812 |
+
825,
|
| 813 |
+
478
|
| 814 |
+
],
|
| 815 |
+
"page_idx": 7
|
| 816 |
+
},
|
| 817 |
+
{
|
| 818 |
+
"type": "text",
|
| 819 |
+
"text": "oisy and Sparse supervision: We provide noisy distances for only 5 trajectories as supervision ",
|
| 820 |
+
"bbox": [
|
| 821 |
+
186,
|
| 822 |
+
478,
|
| 823 |
+
803,
|
| 824 |
+
492
|
| 825 |
+
],
|
| 826 |
+
"page_idx": 7
|
| 827 |
+
},
|
| 828 |
+
{
|
| 829 |
+
"type": "text",
|
| 830 |
+
"text": "Figure 6 shows an example of noisy and sparse supervision. The results are shown in Table 5. The SPT model maintains performance benefits over the baselines under all the settings. Interestingly, under sparse supervision, the map prediction accuracy drops, but the planning accuracy does not drop as much. This is because the model learns to predict minimum map required to predict the action distances of all valid locations accurately as seen in examples shown in Figure 15 in the Appendix. ",
|
| 831 |
+
"bbox": [
|
| 832 |
+
174,
|
| 833 |
+
500,
|
| 834 |
+
825,
|
| 835 |
+
569
|
| 836 |
+
],
|
| 837 |
+
"page_idx": 7
|
| 838 |
+
},
|
| 839 |
+
{
|
| 840 |
+
"type": "text",
|
| 841 |
+
"text": "6 RELATED WORK ",
|
| 842 |
+
"text_level": 1,
|
| 843 |
+
"bbox": [
|
| 844 |
+
176,
|
| 845 |
+
587,
|
| 846 |
+
344,
|
| 847 |
+
603
|
| 848 |
+
],
|
| 849 |
+
"page_idx": 7
|
| 850 |
+
},
|
| 851 |
+
{
|
| 852 |
+
"type": "text",
|
| 853 |
+
"text": "Path planning in known or inferred maps, also known as motion planning in robotics, is a well explored problem led by the seminal papers (Canny, 1988; Kavraki et al., 1996; LaValle & Kuffner Jr, 2001; Karaman & Frazzoli, 2011). Although there are learned variants of motion planners proposed in the literature using gaussian processes (Ijspeert et al., 2013; Ratliff et al., 2018), data-driven motion planners using neural networks is a recent direction (Qureshi et al., 2019; Bhardwaj et al., 2020; Qureshi et al., 2020). Prior work has also studied the use of neural networks to learn the heuristics and sampling stratergies in classical planners Ichter et al. (2018); Guez et al. (2018); Satorras & Welling (2020); Khan et al. (2020). Learning for planning is more common in Markov Decision Process (MDPs) for computing value function via dynamic programming based value iterations (Bellman, 1966; Bertsekas et al., 1995). Planning and learning in neural networks has been explored (Ilin et al., 2007) with a successful general formulation provided by value iteration networks (VIN) (Tamar et al., 2016) with follow-ups to improve scalability and efficiency (Lee et al., 2018; Karkus et al., 2017; Nardelli et al., 2018; Schleich et al., 2019; Khan et al., 2018; Chen et al., 2020). However, these models only capture local value propagation using CNNs and are mostly applied in navigation setups. In contrast, proposed SPTs capture long-range spatial dependency and easily scale to both navigation and manipulation. ",
|
| 854 |
+
"bbox": [
|
| 855 |
+
173,
|
| 856 |
+
612,
|
| 857 |
+
825,
|
| 858 |
+
833
|
| 859 |
+
],
|
| 860 |
+
"page_idx": 7
|
| 861 |
+
},
|
| 862 |
+
{
|
| 863 |
+
"type": "text",
|
| 864 |
+
"text": "Differentiable planning structure has also been explored in reinforcement learning with model-free methods (Silver et al., 2017; Oh et al., 2017; Zhu et al., 2017; Farquhar et al., 2018) as well as off-policy RL (Eysenbach et al., 2019; Laskin et al., 2020). Recent works also backpropagate through learned planners to train the policy (Pathak et al., 2018; Srinivas et al., 2018; Amos et al., 2018) and use imagined rollouts of a learned world model for long-term plans (Racaniere et al. \\` , 2017; Hafner et al., 2019; Sekar et al., 2019). Unlike our work, these works lack the structure of a spatial planner. ",
|
| 865 |
+
"bbox": [
|
| 866 |
+
174,
|
| 867 |
+
840,
|
| 868 |
+
825,
|
| 869 |
+
922
|
| 870 |
+
],
|
| 871 |
+
"page_idx": 7
|
| 872 |
+
},
|
| 873 |
+
{
|
| 874 |
+
"type": "image",
|
| 875 |
+
"img_path": "images/1d3ced193524c5cb8ec9da7382146b7aa337bebf0a5db11eb04e1170421f6acf.jpg",
|
| 876 |
+
"image_caption": [
|
| 877 |
+
"Figure 6: Sparse and Noisy Supervision. Figure showing examples of a map and goal with different levels of supervision. Noisy supervision adds gaussian noise to the ground truth distance values, and sparse supervision samples 5 trajectories for random starting locations to the goal location. "
|
| 878 |
+
],
|
| 879 |
+
"image_footnote": [],
|
| 880 |
+
"bbox": [
|
| 881 |
+
179,
|
| 882 |
+
102,
|
| 883 |
+
818,
|
| 884 |
+
219
|
| 885 |
+
],
|
| 886 |
+
"page_idx": 8
|
| 887 |
+
},
|
| 888 |
+
{
|
| 889 |
+
"type": "table",
|
| 890 |
+
"img_path": "images/a858ebba92465966468fc88d0809c1d41542ea974683494e807de56c07bbdf65.jpg",
|
| 891 |
+
"table_caption": [],
|
| 892 |
+
"table_footnote": [
|
| 893 |
+
"Table 5: Sparse and Noisy Supervision Results. Table showing the average mapping and planning accuracy of the proposed model Spatial Planning Transformer (SPT) as compared to the baselines for end-to-end mapping and planning experiments under noisy and sparse supervision settings for the manipulation task. "
|
| 894 |
+
],
|
| 895 |
+
"table_body": "<table><tr><td></td><td colspan=\"2\">Dense and Perfect supervision</td><td colspan=\"2\">Noisy supervision</td><td colspan=\"2\">Sparse supervision</td><td colspan=\"2\">Noisy and Sparse supervision</td></tr><tr><td>Method</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td><td>Map Acc</td><td>Plan Acc</td></tr><tr><td>VIN</td><td>81.25</td><td>66.45</td><td>75.68</td><td>60.78</td><td>70.16</td><td>60.23</td><td>70.22</td><td>58.97</td></tr><tr><td>GPPN</td><td>85.57</td><td>82.13</td><td>80.13</td><td>76.11</td><td>72.73</td><td>75.13</td><td>70.08</td><td>72.85</td></tr><tr><td>SPT</td><td>98.96</td><td>98.42</td><td>96.35</td><td>95.83</td><td>80.15</td><td>97.18</td><td>77.17</td><td>94.34</td></tr></table>",
|
| 896 |
+
"bbox": [
|
| 897 |
+
173,
|
| 898 |
+
280,
|
| 899 |
+
825,
|
| 900 |
+
386
|
| 901 |
+
],
|
| 902 |
+
"page_idx": 8
|
| 903 |
+
},
|
| 904 |
+
{
|
| 905 |
+
"type": "text",
|
| 906 |
+
"text": "Decomposing learning a controller into mapping and planning is common in robot navigation (Khatib, 1986; Elfes, 1987). Some works have explored joint mapping and planning (Elfes, 1989; Fraundorfer et al., 2012). Maps can also be built from vision (Konolige et al., 2010; Fuentes-Pacheco et al., 2015) with a learned mapper (Parisotto & Salakhutdinov, 2017; Karkus et al., 2020). There has been some work on learning maps without using map annotations as well (Gregor et al., 2019). For navigation specific applications, recent works proposed joint mapping and planning for navigation (Gupta et al., 2017; Zhang et al., 2017; Savinov et al., 2018; Chaplot et al., 2020). However, most of these works either require access to ground truth map or assume interaction. Hence, they will first need to be trained in simulation. In contrast, we show results when the map is not known to the agent by learning just from trajectories and can be directly learned from data collected in the real-world. ",
|
| 907 |
+
"bbox": [
|
| 908 |
+
173,
|
| 909 |
+
455,
|
| 910 |
+
825,
|
| 911 |
+
594
|
| 912 |
+
],
|
| 913 |
+
"page_idx": 8
|
| 914 |
+
},
|
| 915 |
+
{
|
| 916 |
+
"type": "text",
|
| 917 |
+
"text": "7 DISCUSSION ",
|
| 918 |
+
"text_level": 1,
|
| 919 |
+
"bbox": [
|
| 920 |
+
174,
|
| 921 |
+
609,
|
| 922 |
+
310,
|
| 923 |
+
625
|
| 924 |
+
],
|
| 925 |
+
"page_idx": 8
|
| 926 |
+
},
|
| 927 |
+
{
|
| 928 |
+
"type": "text",
|
| 929 |
+
"text": "The SPT model is designed to learn long-range spatial planning and it outperforms baselines across multiple experimental settings on both navigation and manipulation tasks. End-to-end learning experiments demonstrate that the SPT model can deal with unknown maps by learning mapping without any map-supervision. However, there are some limitations that need to be addressed before learning-based planning methods can be used for large-scale applications. We showed that the SPT model scales much better with increasing map sizes as compared to the baselines, however larger map sizes lead to a high inference time. In the future, we plan to tackle larger map sizes by learning an encoding which reduces the size of the map before spatial planning. The action space was limited to 4 actions in our experiments. Actions can be more fine-grained with smaller cells in the map, which can also be incorporated by tackling larger map sizes in the future. ",
|
| 930 |
+
"bbox": [
|
| 931 |
+
174,
|
| 932 |
+
632,
|
| 933 |
+
825,
|
| 934 |
+
772
|
| 935 |
+
],
|
| 936 |
+
"page_idx": 8
|
| 937 |
+
},
|
| 938 |
+
{
|
| 939 |
+
"type": "text",
|
| 940 |
+
"text": "REFERENCES ",
|
| 941 |
+
"text_level": 1,
|
| 942 |
+
"bbox": [
|
| 943 |
+
176,
|
| 944 |
+
794,
|
| 945 |
+
285,
|
| 946 |
+
809
|
| 947 |
+
],
|
| 948 |
+
"page_idx": 8
|
| 949 |
+
},
|
| 950 |
+
{
|
| 951 |
+
"type": "text",
|
| 952 |
+
"text": "Brandon Amos, Ivan Jimenez, Jacob Sacks, Byron Boots, and J Zico Kolter. Differentiable mpc for end-to-end planning and control. In NeurIPS, 2018. ",
|
| 953 |
+
"bbox": [
|
| 954 |
+
174,
|
| 955 |
+
827,
|
| 956 |
+
825,
|
| 957 |
+
856
|
| 958 |
+
],
|
| 959 |
+
"page_idx": 8
|
| 960 |
+
},
|
| 961 |
+
{
|
| 962 |
+
"type": "text",
|
| 963 |
+
"text": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. ",
|
| 964 |
+
"bbox": [
|
| 965 |
+
171,
|
| 966 |
+
867,
|
| 967 |
+
825,
|
| 968 |
+
896
|
| 969 |
+
],
|
| 970 |
+
"page_idx": 8
|
| 971 |
+
},
|
| 972 |
+
{
|
| 973 |
+
"type": "text",
|
| 974 |
+
"text": "Richard Bellman. Dynamic programming. Science, 1966. ",
|
| 975 |
+
"bbox": [
|
| 976 |
+
174,
|
| 977 |
+
909,
|
| 978 |
+
553,
|
| 979 |
+
924
|
| 980 |
+
],
|
| 981 |
+
"page_idx": 8
|
| 982 |
+
},
|
| 983 |
+
{
|
| 984 |
+
"type": "text",
|
| 985 |
+
"text": "Dimitri P Bertsekas, Dimitri P Bertsekas, Dimitri P Bertsekas, and Dimitri P Bertsekas. Dynamic programming and optimal control. Athena scientific Belmont, MA, 1995. ",
|
| 986 |
+
"bbox": [
|
| 987 |
+
171,
|
| 988 |
+
103,
|
| 989 |
+
823,
|
| 990 |
+
132
|
| 991 |
+
],
|
| 992 |
+
"page_idx": 9
|
| 993 |
+
},
|
| 994 |
+
{
|
| 995 |
+
"type": "text",
|
| 996 |
+
"text": "Mohak Bhardwaj, Byron Boots, and Mustafa Mukadam. Differentiable gaussian process motion planning. In International Conference on Robotics and Automation (ICRA), 2020. ",
|
| 997 |
+
"bbox": [
|
| 998 |
+
171,
|
| 999 |
+
141,
|
| 1000 |
+
823,
|
| 1001 |
+
171
|
| 1002 |
+
],
|
| 1003 |
+
"page_idx": 9
|
| 1004 |
+
},
|
| 1005 |
+
{
|
| 1006 |
+
"type": "text",
|
| 1007 |
+
"text": "Leon Bottou. Large-scale machine learning with stochastic gradient descent. In ´ Proceedings of COMPSTAT’2010. Springer, 2010. ",
|
| 1008 |
+
"bbox": [
|
| 1009 |
+
173,
|
| 1010 |
+
179,
|
| 1011 |
+
825,
|
| 1012 |
+
209
|
| 1013 |
+
],
|
| 1014 |
+
"page_idx": 9
|
| 1015 |
+
},
|
| 1016 |
+
{
|
| 1017 |
+
"type": "text",
|
| 1018 |
+
"text": "John Canny. The complexity of robot motion planning. MIT press, 1988. ",
|
| 1019 |
+
"bbox": [
|
| 1020 |
+
174,
|
| 1021 |
+
217,
|
| 1022 |
+
650,
|
| 1023 |
+
233
|
| 1024 |
+
],
|
| 1025 |
+
"page_idx": 9
|
| 1026 |
+
},
|
| 1027 |
+
{
|
| 1028 |
+
"type": "text",
|
| 1029 |
+
"text": "Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. arXiv preprint arXiv:2004.05155, 2020. ",
|
| 1030 |
+
"bbox": [
|
| 1031 |
+
174,
|
| 1032 |
+
242,
|
| 1033 |
+
826,
|
| 1034 |
+
271
|
| 1035 |
+
],
|
| 1036 |
+
"page_idx": 9
|
| 1037 |
+
},
|
| 1038 |
+
{
|
| 1039 |
+
"type": "text",
|
| 1040 |
+
"text": "Binghong Chen, Bo Dai, Qinjie Lin, Guo Ye, Han Liu, and Le Song. Learning to plan in high dimensions via neural exploration-exploitation trees. In ICLR, 2020. ",
|
| 1041 |
+
"bbox": [
|
| 1042 |
+
173,
|
| 1043 |
+
280,
|
| 1044 |
+
825,
|
| 1045 |
+
309
|
| 1046 |
+
],
|
| 1047 |
+
"page_idx": 9
|
| 1048 |
+
},
|
| 1049 |
+
{
|
| 1050 |
+
"type": "text",
|
| 1051 |
+
"text": "Edsger W Dijkstra et al. A note on two problems in connexion with graphs. Numerische mathematik, 1959. ",
|
| 1052 |
+
"bbox": [
|
| 1053 |
+
171,
|
| 1054 |
+
318,
|
| 1055 |
+
826,
|
| 1056 |
+
347
|
| 1057 |
+
],
|
| 1058 |
+
"page_idx": 9
|
| 1059 |
+
},
|
| 1060 |
+
{
|
| 1061 |
+
"type": "text",
|
| 1062 |
+
"text": "Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 2006. ",
|
| 1063 |
+
"bbox": [
|
| 1064 |
+
173,
|
| 1065 |
+
356,
|
| 1066 |
+
823,
|
| 1067 |
+
386
|
| 1068 |
+
],
|
| 1069 |
+
"page_idx": 9
|
| 1070 |
+
},
|
| 1071 |
+
{
|
| 1072 |
+
"type": "text",
|
| 1073 |
+
"text": "Alberto Elfes. Sonar-based real-world mapping and navigation. IEEE Journal on Robotics and Automation, 1987. ",
|
| 1074 |
+
"bbox": [
|
| 1075 |
+
173,
|
| 1076 |
+
395,
|
| 1077 |
+
823,
|
| 1078 |
+
424
|
| 1079 |
+
],
|
| 1080 |
+
"page_idx": 9
|
| 1081 |
+
},
|
| 1082 |
+
{
|
| 1083 |
+
"type": "text",
|
| 1084 |
+
"text": "Alberto Elfes. Using occupancy grids for mobile robot perception and navigation. Computer, 1989. ",
|
| 1085 |
+
"bbox": [
|
| 1086 |
+
173,
|
| 1087 |
+
433,
|
| 1088 |
+
823,
|
| 1089 |
+
449
|
| 1090 |
+
],
|
| 1091 |
+
"page_idx": 9
|
| 1092 |
+
},
|
| 1093 |
+
{
|
| 1094 |
+
"type": "text",
|
| 1095 |
+
"text": "Ben Eysenbach, Russ R Salakhutdinov, and Sergey Levine. Search on the replay buffer: Bridging planning and reinforcement learning. In NeurIPS, 2019. ",
|
| 1096 |
+
"bbox": [
|
| 1097 |
+
174,
|
| 1098 |
+
457,
|
| 1099 |
+
823,
|
| 1100 |
+
487
|
| 1101 |
+
],
|
| 1102 |
+
"page_idx": 9
|
| 1103 |
+
},
|
| 1104 |
+
{
|
| 1105 |
+
"type": "text",
|
| 1106 |
+
"text": "Gregory Farquhar, Tim Rocktaschel, Maximilian Igl, and Shimon Whiteson. Treeqn and atreec: ¨ Differentiable tree-structured models for deep reinforcement learning. In ICLR, 2018. ",
|
| 1107 |
+
"bbox": [
|
| 1108 |
+
174,
|
| 1109 |
+
496,
|
| 1110 |
+
823,
|
| 1111 |
+
525
|
| 1112 |
+
],
|
| 1113 |
+
"page_idx": 9
|
| 1114 |
+
},
|
| 1115 |
+
{
|
| 1116 |
+
"type": "text",
|
| 1117 |
+
"text": "Friedrich Fraundorfer, Lionel Heng, Dominik Honegger, Gim Hee Lee, Lorenz Meier, Petri Tanskanen, and Marc Pollefeys. Vision-based autonomous mapping and exploration using a quadrotor mav. In International Conference on Intelligent Robots and Systems, 2012. ",
|
| 1118 |
+
"bbox": [
|
| 1119 |
+
174,
|
| 1120 |
+
534,
|
| 1121 |
+
826,
|
| 1122 |
+
577
|
| 1123 |
+
],
|
| 1124 |
+
"page_idx": 9
|
| 1125 |
+
},
|
| 1126 |
+
{
|
| 1127 |
+
"type": "text",
|
| 1128 |
+
"text": "Jorge Fuentes-Pacheco, Jose Ruiz-Ascencio, and Juan Manuel Rend ´ on-Mancha. Visual simultaneous ´ localization and mapping: a survey. Artificial intelligence review, 2015. ",
|
| 1129 |
+
"bbox": [
|
| 1130 |
+
174,
|
| 1131 |
+
585,
|
| 1132 |
+
823,
|
| 1133 |
+
616
|
| 1134 |
+
],
|
| 1135 |
+
"page_idx": 9
|
| 1136 |
+
},
|
| 1137 |
+
{
|
| 1138 |
+
"type": "text",
|
| 1139 |
+
"text": "Tobias Glasmachers. Limits of end-to-end learning. arXiv preprint arXiv:1704.08305, 2017. ",
|
| 1140 |
+
"bbox": [
|
| 1141 |
+
171,
|
| 1142 |
+
623,
|
| 1143 |
+
782,
|
| 1144 |
+
640
|
| 1145 |
+
],
|
| 1146 |
+
"page_idx": 9
|
| 1147 |
+
},
|
| 1148 |
+
{
|
| 1149 |
+
"type": "text",
|
| 1150 |
+
"text": "Reginald G Golledge, Valerie Dougherty, and Scott Bell. Acquiring spatial knowledge: Survey versus route-based knowledge in unfamiliar environments. Annals of the association of American geographers, 1995. ",
|
| 1151 |
+
"bbox": [
|
| 1152 |
+
174,
|
| 1153 |
+
648,
|
| 1154 |
+
826,
|
| 1155 |
+
691
|
| 1156 |
+
],
|
| 1157 |
+
"page_idx": 9
|
| 1158 |
+
},
|
| 1159 |
+
{
|
| 1160 |
+
"type": "text",
|
| 1161 |
+
"text": "Karol Gregor, Danilo Jimenez Rezende, Frederic Besse, Yan Wu, Hamza Merzic, and Aaron van den Oord. Shaping belief states with generative environment models for rl. Advances in Neural Information Processing Systems, 32:13475–13487, 2019. ",
|
| 1162 |
+
"bbox": [
|
| 1163 |
+
174,
|
| 1164 |
+
700,
|
| 1165 |
+
825,
|
| 1166 |
+
743
|
| 1167 |
+
],
|
| 1168 |
+
"page_idx": 9
|
| 1169 |
+
},
|
| 1170 |
+
{
|
| 1171 |
+
"type": "text",
|
| 1172 |
+
"text": "Arthur Guez, Theophane Weber, Ioannis Antonoglou, Karen Simonyan, Oriol Vinyals, Daan Wierstra, Remi Munos, and David Silver. Learning to search with mctsnets. In International Conference on Machine Learning, pp. 1822–1831, 2018. ",
|
| 1173 |
+
"bbox": [
|
| 1174 |
+
174,
|
| 1175 |
+
752,
|
| 1176 |
+
825,
|
| 1177 |
+
796
|
| 1178 |
+
],
|
| 1179 |
+
"page_idx": 9
|
| 1180 |
+
},
|
| 1181 |
+
{
|
| 1182 |
+
"type": "text",
|
| 1183 |
+
"text": "Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. ",
|
| 1184 |
+
"bbox": [
|
| 1185 |
+
174,
|
| 1186 |
+
804,
|
| 1187 |
+
823,
|
| 1188 |
+
848
|
| 1189 |
+
],
|
| 1190 |
+
"page_idx": 9
|
| 1191 |
+
},
|
| 1192 |
+
{
|
| 1193 |
+
"type": "text",
|
| 1194 |
+
"text": "Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In ICML, 2019. ",
|
| 1195 |
+
"bbox": [
|
| 1196 |
+
174,
|
| 1197 |
+
857,
|
| 1198 |
+
821,
|
| 1199 |
+
886
|
| 1200 |
+
],
|
| 1201 |
+
"page_idx": 9
|
| 1202 |
+
},
|
| 1203 |
+
{
|
| 1204 |
+
"type": "text",
|
| 1205 |
+
"text": "Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100–107, 1968. ",
|
| 1206 |
+
"bbox": [
|
| 1207 |
+
174,
|
| 1208 |
+
895,
|
| 1209 |
+
823,
|
| 1210 |
+
924
|
| 1211 |
+
],
|
| 1212 |
+
"page_idx": 9
|
| 1213 |
+
},
|
| 1214 |
+
{
|
| 1215 |
+
"type": "text",
|
| 1216 |
+
"text": "Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 1043–1051. IEEE, 2019. ",
|
| 1217 |
+
"bbox": [
|
| 1218 |
+
176,
|
| 1219 |
+
103,
|
| 1220 |
+
823,
|
| 1221 |
+
146
|
| 1222 |
+
],
|
| 1223 |
+
"page_idx": 10
|
| 1224 |
+
},
|
| 1225 |
+
{
|
| 1226 |
+
"type": "text",
|
| 1227 |
+
"text": "Brian Ichter, James Harrison, and Marco Pavone. Learning sampling distributions for robot motion planning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 7087–7094. IEEE, 2018. ",
|
| 1228 |
+
"bbox": [
|
| 1229 |
+
178,
|
| 1230 |
+
155,
|
| 1231 |
+
823,
|
| 1232 |
+
198
|
| 1233 |
+
],
|
| 1234 |
+
"page_idx": 10
|
| 1235 |
+
},
|
| 1236 |
+
{
|
| 1237 |
+
"type": "text",
|
| 1238 |
+
"text": "Auke Jan Ijspeert, Jun Nakanishi, Heiko Hoffmann, Peter Pastor, and Stefan Schaal. Dynamical movement primitives: learning attractor models for motor behaviors. Neural computation, 2013. ",
|
| 1239 |
+
"bbox": [
|
| 1240 |
+
174,
|
| 1241 |
+
207,
|
| 1242 |
+
823,
|
| 1243 |
+
236
|
| 1244 |
+
],
|
| 1245 |
+
"page_idx": 10
|
| 1246 |
+
},
|
| 1247 |
+
{
|
| 1248 |
+
"type": "text",
|
| 1249 |
+
"text": "Roman Ilin, Robert Kozma, and Paul J Werbos. Efficient learning in cellular simultaneous recurrent neural networks-the case of maze navigation problem. In 2007 IEEE International Symposium on Approximate Dynamic Programming and Reinforcement Learning, 2007. ",
|
| 1250 |
+
"bbox": [
|
| 1251 |
+
173,
|
| 1252 |
+
244,
|
| 1253 |
+
826,
|
| 1254 |
+
287
|
| 1255 |
+
],
|
| 1256 |
+
"page_idx": 10
|
| 1257 |
+
},
|
| 1258 |
+
{
|
| 1259 |
+
"type": "text",
|
| 1260 |
+
"text": "Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 2011. ",
|
| 1261 |
+
"bbox": [
|
| 1262 |
+
174,
|
| 1263 |
+
296,
|
| 1264 |
+
823,
|
| 1265 |
+
327
|
| 1266 |
+
],
|
| 1267 |
+
"page_idx": 10
|
| 1268 |
+
},
|
| 1269 |
+
{
|
| 1270 |
+
"type": "text",
|
| 1271 |
+
"text": "Peter Karkus, David Hsu, and Wee Sun Lee. Qmdp-net: Deep learning for planning under partial observability. In Advances in Neural Information Processing Systems, 2017. ",
|
| 1272 |
+
"bbox": [
|
| 1273 |
+
173,
|
| 1274 |
+
334,
|
| 1275 |
+
823,
|
| 1276 |
+
364
|
| 1277 |
+
],
|
| 1278 |
+
"page_idx": 10
|
| 1279 |
+
},
|
| 1280 |
+
{
|
| 1281 |
+
"type": "text",
|
| 1282 |
+
"text": "Peter Karkus, Anelia Angelova, Vincent Vanhoucke, and Rico Jonschkowski. Differentiable mapping networks: Learning structured map representations for sparse visual localization. arXiv preprint arXiv:2005.09530, 2020. ",
|
| 1283 |
+
"bbox": [
|
| 1284 |
+
173,
|
| 1285 |
+
372,
|
| 1286 |
+
825,
|
| 1287 |
+
415
|
| 1288 |
+
],
|
| 1289 |
+
"page_idx": 10
|
| 1290 |
+
},
|
| 1291 |
+
{
|
| 1292 |
+
"type": "text",
|
| 1293 |
+
"text": "Lydia E Kavraki, Petr Svestka, J-C Latombe, and Mark H Overmars. Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE transactions on Robotics and Automation, 1996. ",
|
| 1294 |
+
"bbox": [
|
| 1295 |
+
173,
|
| 1296 |
+
424,
|
| 1297 |
+
825,
|
| 1298 |
+
467
|
| 1299 |
+
],
|
| 1300 |
+
"page_idx": 10
|
| 1301 |
+
},
|
| 1302 |
+
{
|
| 1303 |
+
"type": "text",
|
| 1304 |
+
"text": "Arbaaz Khan, Clark Zhang, Nikolay Atanasov, Konstantinos Karydis, Vijay Kumar, and Daniel D Lee. Memory augmented control networks. In ICLR, 2018. ",
|
| 1305 |
+
"bbox": [
|
| 1306 |
+
171,
|
| 1307 |
+
477,
|
| 1308 |
+
823,
|
| 1309 |
+
506
|
| 1310 |
+
],
|
| 1311 |
+
"page_idx": 10
|
| 1312 |
+
},
|
| 1313 |
+
{
|
| 1314 |
+
"type": "text",
|
| 1315 |
+
"text": "Arbaaz Khan, Alejandro Ribeiro, Vijay Kumar, and Anthony G Francis. Graph neural networks for motion planning. arXiv preprint arXiv:2006.06248, 2020. ",
|
| 1316 |
+
"bbox": [
|
| 1317 |
+
171,
|
| 1318 |
+
515,
|
| 1319 |
+
823,
|
| 1320 |
+
544
|
| 1321 |
+
],
|
| 1322 |
+
"page_idx": 10
|
| 1323 |
+
},
|
| 1324 |
+
{
|
| 1325 |
+
"type": "text",
|
| 1326 |
+
"text": "Oussama Khatib. Real-time obstacle avoidance for manipulators and mobile robots. In Autonomous robot vehicles. Springer, 1986. ",
|
| 1327 |
+
"bbox": [
|
| 1328 |
+
173,
|
| 1329 |
+
553,
|
| 1330 |
+
823,
|
| 1331 |
+
582
|
| 1332 |
+
],
|
| 1333 |
+
"page_idx": 10
|
| 1334 |
+
},
|
| 1335 |
+
{
|
| 1336 |
+
"type": "text",
|
| 1337 |
+
"text": "Kurt Konolige, James Bowman, JD Chen, Patrick Mihelich, Michael Calonder, Vincent Lepetit, and Pascal Fua. View-based maps. The International Journal of Robotics Research, 2010. ",
|
| 1338 |
+
"bbox": [
|
| 1339 |
+
173,
|
| 1340 |
+
590,
|
| 1341 |
+
823,
|
| 1342 |
+
621
|
| 1343 |
+
],
|
| 1344 |
+
"page_idx": 10
|
| 1345 |
+
},
|
| 1346 |
+
{
|
| 1347 |
+
"type": "text",
|
| 1348 |
+
"text": "Michael Laskin, Scott Emmons, Ajay Jain, Thanard Kurutach, Pieter Abbeel, and Deepak Pathak. Sparse graphical memory for robust planning. arXiv preprint arXiv:2003.06417, 2020. ",
|
| 1349 |
+
"bbox": [
|
| 1350 |
+
171,
|
| 1351 |
+
628,
|
| 1352 |
+
825,
|
| 1353 |
+
659
|
| 1354 |
+
],
|
| 1355 |
+
"page_idx": 10
|
| 1356 |
+
},
|
| 1357 |
+
{
|
| 1358 |
+
"type": "text",
|
| 1359 |
+
"text": "Steven M LaValle and James J Kuffner Jr. Randomized kinodynamic planning. The international journal of robotics research, 2001. ",
|
| 1360 |
+
"bbox": [
|
| 1361 |
+
171,
|
| 1362 |
+
666,
|
| 1363 |
+
825,
|
| 1364 |
+
695
|
| 1365 |
+
],
|
| 1366 |
+
"page_idx": 10
|
| 1367 |
+
},
|
| 1368 |
+
{
|
| 1369 |
+
"type": "text",
|
| 1370 |
+
"text": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 1998. ",
|
| 1371 |
+
"bbox": [
|
| 1372 |
+
171,
|
| 1373 |
+
704,
|
| 1374 |
+
823,
|
| 1375 |
+
734
|
| 1376 |
+
],
|
| 1377 |
+
"page_idx": 10
|
| 1378 |
+
},
|
| 1379 |
+
{
|
| 1380 |
+
"type": "text",
|
| 1381 |
+
"text": "Lisa Lee, Emilio Parisotto, Devendra Singh Chaplot, Eric Xing, and Ruslan Salakhutdinov. Gated path planning networks. In ICML, 2018. ",
|
| 1382 |
+
"bbox": [
|
| 1383 |
+
171,
|
| 1384 |
+
742,
|
| 1385 |
+
825,
|
| 1386 |
+
772
|
| 1387 |
+
],
|
| 1388 |
+
"page_idx": 10
|
| 1389 |
+
},
|
| 1390 |
+
{
|
| 1391 |
+
"type": "text",
|
| 1392 |
+
"text": "Tomas Lozano-Perez. Spatial planning: A configuration space approach. In Autonomous robot vehicles. Springer, 1990. ",
|
| 1393 |
+
"bbox": [
|
| 1394 |
+
173,
|
| 1395 |
+
781,
|
| 1396 |
+
823,
|
| 1397 |
+
810
|
| 1398 |
+
],
|
| 1399 |
+
"page_idx": 10
|
| 1400 |
+
},
|
| 1401 |
+
{
|
| 1402 |
+
"type": "text",
|
| 1403 |
+
"text": "Dmytro Mishkin, Alexey Dosovitskiy, and Vladlen Koltun. Benchmarking classic and learned navigation in complex 3d environments. arXiv preprint arXiv:1901.10915, 2019. ",
|
| 1404 |
+
"bbox": [
|
| 1405 |
+
174,
|
| 1406 |
+
819,
|
| 1407 |
+
823,
|
| 1408 |
+
848
|
| 1409 |
+
],
|
| 1410 |
+
"page_idx": 10
|
| 1411 |
+
},
|
| 1412 |
+
{
|
| 1413 |
+
"type": "text",
|
| 1414 |
+
"text": "Nantas Nardelli, Gabriel Synnaeve, Zeming Lin, Pushmeet Kohli, Philip HS Torr, and Nicolas Usunier. Value propagation networks. arXiv preprint arXiv:1805.11199, 2018. ",
|
| 1415 |
+
"bbox": [
|
| 1416 |
+
176,
|
| 1417 |
+
857,
|
| 1418 |
+
821,
|
| 1419 |
+
886
|
| 1420 |
+
],
|
| 1421 |
+
"page_idx": 10
|
| 1422 |
+
},
|
| 1423 |
+
{
|
| 1424 |
+
"type": "text",
|
| 1425 |
+
"text": "Junhyuk Oh, Satinder Singh, and Honglak Lee. Value prediction network. In Advances in Neural Information Processing Systems, 2017. ",
|
| 1426 |
+
"bbox": [
|
| 1427 |
+
174,
|
| 1428 |
+
895,
|
| 1429 |
+
821,
|
| 1430 |
+
924
|
| 1431 |
+
],
|
| 1432 |
+
"page_idx": 10
|
| 1433 |
+
},
|
| 1434 |
+
{
|
| 1435 |
+
"type": "text",
|
| 1436 |
+
"text": "Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. arXiv preprint arXiv:1702.08360, 2017. ",
|
| 1437 |
+
"bbox": [
|
| 1438 |
+
171,
|
| 1439 |
+
103,
|
| 1440 |
+
823,
|
| 1441 |
+
132
|
| 1442 |
+
],
|
| 1443 |
+
"page_idx": 11
|
| 1444 |
+
},
|
| 1445 |
+
{
|
| 1446 |
+
"type": "text",
|
| 1447 |
+
"text": "Deepak Pathak, Parsa Mahmoudieh, Guanghao Luo, Pulkit Agrawal, Dian Chen, Yide Shentu, Evan Shelhamer, Jitendra Malik, Alexei A Efros, and Trevor Darrell. Zero-shot visual imitation. In ICLR, 2018. ",
|
| 1448 |
+
"bbox": [
|
| 1449 |
+
174,
|
| 1450 |
+
140,
|
| 1451 |
+
823,
|
| 1452 |
+
184
|
| 1453 |
+
],
|
| 1454 |
+
"page_idx": 11
|
| 1455 |
+
},
|
| 1456 |
+
{
|
| 1457 |
+
"type": "text",
|
| 1458 |
+
"text": "Ahmed H Qureshi, Anthony Simeonov, Mayur J Bency, and Michael C Yip. Motion planning networks. In International Conference on Robotics and Automation (ICRA), 2019. ",
|
| 1459 |
+
"bbox": [
|
| 1460 |
+
171,
|
| 1461 |
+
193,
|
| 1462 |
+
823,
|
| 1463 |
+
222
|
| 1464 |
+
],
|
| 1465 |
+
"page_idx": 11
|
| 1466 |
+
},
|
| 1467 |
+
{
|
| 1468 |
+
"type": "text",
|
| 1469 |
+
"text": "Ahmed H Qureshi, Jiangeng Dong, Austin Choe, and Michael C Yip. Neural manipulation planning on constraint manifolds. IEEE Robotics and Automation Letters, 5(4):6089–6096, 2020. ",
|
| 1470 |
+
"bbox": [
|
| 1471 |
+
173,
|
| 1472 |
+
229,
|
| 1473 |
+
823,
|
| 1474 |
+
260
|
| 1475 |
+
],
|
| 1476 |
+
"page_idx": 11
|
| 1477 |
+
},
|
| 1478 |
+
{
|
| 1479 |
+
"type": "text",
|
| 1480 |
+
"text": "Sebastien Racani ´ ere, Th \\` eophane Weber, David Reichert, Lars Buesing, Arthur Guez, Danilo Jimenez ´ Rezende, Adria Puigdomenech Badia, Oriol Vinyals, Nicolas Heess, Yujia Li, et al. Imaginationaugmented agents for deep reinforcement learning. In NIPS, 2017. ",
|
| 1481 |
+
"bbox": [
|
| 1482 |
+
174,
|
| 1483 |
+
267,
|
| 1484 |
+
823,
|
| 1485 |
+
310
|
| 1486 |
+
],
|
| 1487 |
+
"page_idx": 11
|
| 1488 |
+
},
|
| 1489 |
+
{
|
| 1490 |
+
"type": "text",
|
| 1491 |
+
"text": "Nathan D Ratliff, Jan Issac, Daniel Kappler, Stan Birchfield, and Dieter Fox. Riemannian motion policies. arXiv preprint arXiv:1801.02854, 2018. ",
|
| 1492 |
+
"bbox": [
|
| 1493 |
+
173,
|
| 1494 |
+
319,
|
| 1495 |
+
821,
|
| 1496 |
+
349
|
| 1497 |
+
],
|
| 1498 |
+
"page_idx": 11
|
| 1499 |
+
},
|
| 1500 |
+
{
|
| 1501 |
+
"type": "text",
|
| 1502 |
+
"text": "Victor Garcia Satorras and Max Welling. Neural enhanced belief propagation on factor graphs. arXiv preprint arXiv:2003.01998, 2020. ",
|
| 1503 |
+
"bbox": [
|
| 1504 |
+
174,
|
| 1505 |
+
357,
|
| 1506 |
+
821,
|
| 1507 |
+
386
|
| 1508 |
+
],
|
| 1509 |
+
"page_idx": 11
|
| 1510 |
+
},
|
| 1511 |
+
{
|
| 1512 |
+
"type": "text",
|
| 1513 |
+
"text": "Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. In ICLR, 2018. ",
|
| 1514 |
+
"bbox": [
|
| 1515 |
+
173,
|
| 1516 |
+
395,
|
| 1517 |
+
821,
|
| 1518 |
+
424
|
| 1519 |
+
],
|
| 1520 |
+
"page_idx": 11
|
| 1521 |
+
},
|
| 1522 |
+
{
|
| 1523 |
+
"type": "text",
|
| 1524 |
+
"text": "Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. ",
|
| 1525 |
+
"bbox": [
|
| 1526 |
+
173,
|
| 1527 |
+
433,
|
| 1528 |
+
823,
|
| 1529 |
+
476
|
| 1530 |
+
],
|
| 1531 |
+
"page_idx": 11
|
| 1532 |
+
},
|
| 1533 |
+
{
|
| 1534 |
+
"type": "text",
|
| 1535 |
+
"text": "Daniel Schleich, Tobias Klamt, and Sven Behnke. Value iteration networks on multiple levels of abstraction. arXiv preprint arXiv:1905.11068, 2019. ",
|
| 1536 |
+
"bbox": [
|
| 1537 |
+
173,
|
| 1538 |
+
484,
|
| 1539 |
+
823,
|
| 1540 |
+
513
|
| 1541 |
+
],
|
| 1542 |
+
"page_idx": 11
|
| 1543 |
+
},
|
| 1544 |
+
{
|
| 1545 |
+
"type": "text",
|
| 1546 |
+
"text": "Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In ICML, 2019. ",
|
| 1547 |
+
"bbox": [
|
| 1548 |
+
173,
|
| 1549 |
+
522,
|
| 1550 |
+
823,
|
| 1551 |
+
551
|
| 1552 |
+
],
|
| 1553 |
+
"page_idx": 11
|
| 1554 |
+
},
|
| 1555 |
+
{
|
| 1556 |
+
"type": "text",
|
| 1557 |
+
"text": "David Silver, Hado Hasselt, Matteo Hessel, Tom Schaul, Arthur Guez, Tim Harley, Gabriel DulacArnold, David Reichert, Neil Rabinowitz, Andre Barreto, et al. The predictron: End-to-end learning and planning. In International Conference on Machine Learning. PMLR, 2017. ",
|
| 1558 |
+
"bbox": [
|
| 1559 |
+
176,
|
| 1560 |
+
559,
|
| 1561 |
+
825,
|
| 1562 |
+
603
|
| 1563 |
+
],
|
| 1564 |
+
"page_idx": 11
|
| 1565 |
+
},
|
| 1566 |
+
{
|
| 1567 |
+
"type": "text",
|
| 1568 |
+
"text": "Aravind Srinivas, Allan Jabri, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Universal planning networks. ICML, 2018. ",
|
| 1569 |
+
"bbox": [
|
| 1570 |
+
173,
|
| 1571 |
+
611,
|
| 1572 |
+
823,
|
| 1573 |
+
640
|
| 1574 |
+
],
|
| 1575 |
+
"page_idx": 11
|
| 1576 |
+
},
|
| 1577 |
+
{
|
| 1578 |
+
"type": "text",
|
| 1579 |
+
"text": "Aviv Tamar, Yi Wu, Garrett Thomas, Sergey Levine, and Pieter Abbeel. Value iteration networks. In NIPS, 2016. ",
|
| 1580 |
+
"bbox": [
|
| 1581 |
+
173,
|
| 1582 |
+
648,
|
| 1583 |
+
823,
|
| 1584 |
+
678
|
| 1585 |
+
],
|
| 1586 |
+
"page_idx": 11
|
| 1587 |
+
},
|
| 1588 |
+
{
|
| 1589 |
+
"type": "text",
|
| 1590 |
+
"text": "Edward C Tolman. Cognitive maps in rats and men. Psychological review, 1948. ",
|
| 1591 |
+
"bbox": [
|
| 1592 |
+
174,
|
| 1593 |
+
686,
|
| 1594 |
+
707,
|
| 1595 |
+
702
|
| 1596 |
+
],
|
| 1597 |
+
"page_idx": 11
|
| 1598 |
+
},
|
| 1599 |
+
{
|
| 1600 |
+
"type": "text",
|
| 1601 |
+
"text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. ",
|
| 1602 |
+
"bbox": [
|
| 1603 |
+
173,
|
| 1604 |
+
710,
|
| 1605 |
+
825,
|
| 1606 |
+
739
|
| 1607 |
+
],
|
| 1608 |
+
"page_idx": 11
|
| 1609 |
+
},
|
| 1610 |
+
{
|
| 1611 |
+
"type": "text",
|
| 1612 |
+
"text": "Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In CVPR, 2018. ",
|
| 1613 |
+
"bbox": [
|
| 1614 |
+
174,
|
| 1615 |
+
748,
|
| 1616 |
+
823,
|
| 1617 |
+
777
|
| 1618 |
+
],
|
| 1619 |
+
"page_idx": 11
|
| 1620 |
+
},
|
| 1621 |
+
{
|
| 1622 |
+
"type": "text",
|
| 1623 |
+
"text": "Jingwei Zhang, Lei Tai, Joschka Boedecker, Wolfram Burgard, and Ming Liu. Neural slam: Learning to explore with external memory. arXiv preprint arXiv:1706.09520, 2017. ",
|
| 1624 |
+
"bbox": [
|
| 1625 |
+
173,
|
| 1626 |
+
786,
|
| 1627 |
+
823,
|
| 1628 |
+
815
|
| 1629 |
+
],
|
| 1630 |
+
"page_idx": 11
|
| 1631 |
+
},
|
| 1632 |
+
{
|
| 1633 |
+
"type": "text",
|
| 1634 |
+
"text": "Yuke Zhu, Daniel Gordon, Eric Kolve, Dieter Fox, Li Fei-Fei, Abhinav Gupta, Roozbeh Mottaghi, and Ali Farhadi. Visual semantic planning using deep successor representations. In ICCV, 2017. ",
|
| 1635 |
+
"bbox": [
|
| 1636 |
+
173,
|
| 1637 |
+
824,
|
| 1638 |
+
825,
|
| 1639 |
+
853
|
| 1640 |
+
],
|
| 1641 |
+
"page_idx": 11
|
| 1642 |
+
},
|
| 1643 |
+
{
|
| 1644 |
+
"type": "text",
|
| 1645 |
+
"text": "A BACKGROUND: TRANSFORMERS ",
|
| 1646 |
+
"text_level": 1,
|
| 1647 |
+
"bbox": [
|
| 1648 |
+
176,
|
| 1649 |
+
102,
|
| 1650 |
+
482,
|
| 1651 |
+
118
|
| 1652 |
+
],
|
| 1653 |
+
"page_idx": 12
|
| 1654 |
+
},
|
| 1655 |
+
{
|
| 1656 |
+
"type": "text",
|
| 1657 |
+
"text": "The proposed spatial planning method is based on the Transformer model (Vaswani et al., 2017). A Transformer layer, denoted by $f _ { \\mathrm { T L } }$ , takes a tensor $X \\in \\mathcal { R } ^ { d \\times S }$ as input, where $d$ is the embedding size and $S$ is the size of the input. It consists of two sublayers, a multi-head self-attention layer $( f _ { \\mathrm { S A } } )$ and a position-wise fully connected layer $( f _ { \\mathrm { F C } } )$ . There is a residual connection around each sublayer, followed by layer normalization (Ba et al., 2016) (LN): ",
|
| 1658 |
+
"bbox": [
|
| 1659 |
+
173,
|
| 1660 |
+
127,
|
| 1661 |
+
826,
|
| 1662 |
+
198
|
| 1663 |
+
],
|
| 1664 |
+
"page_idx": 12
|
| 1665 |
+
},
|
| 1666 |
+
{
|
| 1667 |
+
"type": "equation",
|
| 1668 |
+
"img_path": "images/e3f1bc7e4437f6f7b2c22d13af66565f77f5afa1745d7f7c4c91f31cc4009de4.jpg",
|
| 1669 |
+
"text": "$$\nR = \\operatorname { L N } ( f _ { \\operatorname { S A } } ( X ) + X ) , \\quad Y = f _ { \\operatorname { T L } } ( X ) = \\operatorname { L N } ( f _ { \\operatorname { F C } } ( R ) + R )\n$$",
|
| 1670 |
+
"text_format": "latex",
|
| 1671 |
+
"bbox": [
|
| 1672 |
+
297,
|
| 1673 |
+
204,
|
| 1674 |
+
700,
|
| 1675 |
+
222
|
| 1676 |
+
],
|
| 1677 |
+
"page_idx": 12
|
| 1678 |
+
},
|
| 1679 |
+
{
|
| 1680 |
+
"type": "text",
|
| 1681 |
+
"text": "where $R , Y \\in \\mathcal { R } ^ { d \\times S }$ are the intermediate and final representations, respectively. ",
|
| 1682 |
+
"bbox": [
|
| 1683 |
+
174,
|
| 1684 |
+
228,
|
| 1685 |
+
696,
|
| 1686 |
+
243
|
| 1687 |
+
],
|
| 1688 |
+
"page_idx": 12
|
| 1689 |
+
},
|
| 1690 |
+
{
|
| 1691 |
+
"type": "text",
|
| 1692 |
+
"text": "The multi-head self-attention $( f _ { \\mathrm { S A } } )$ layer has $h$ attention heads, each computes a scaled dot-product attention over queries $Q$ , keys $K$ and values $V$ , which are all different projections of the input $X$ : ",
|
| 1693 |
+
"bbox": [
|
| 1694 |
+
173,
|
| 1695 |
+
250,
|
| 1696 |
+
825,
|
| 1697 |
+
279
|
| 1698 |
+
],
|
| 1699 |
+
"page_idx": 12
|
| 1700 |
+
},
|
| 1701 |
+
{
|
| 1702 |
+
"type": "equation",
|
| 1703 |
+
"img_path": "images/1698c45a3865a5fa60ed254f4ab1ebfcf219869442dc7bb8556044ba05067269.jpg",
|
| 1704 |
+
"text": "$$\nQ _ { i } = W _ { Q , i } ^ { T } X , \\quad K _ { i } = W _ { K , i } ^ { T } X , \\quad V = W _ { V , i } ^ { T } X\n$$",
|
| 1705 |
+
"text_format": "latex",
|
| 1706 |
+
"bbox": [
|
| 1707 |
+
341,
|
| 1708 |
+
284,
|
| 1709 |
+
656,
|
| 1710 |
+
304
|
| 1711 |
+
],
|
| 1712 |
+
"page_idx": 12
|
| 1713 |
+
},
|
| 1714 |
+
{
|
| 1715 |
+
"type": "equation",
|
| 1716 |
+
"img_path": "images/068c5e069c8b9d12d38fd5e61430b45c46da032089a54844a94ca9974faf8ba4.jpg",
|
| 1717 |
+
"text": "$$\nZ _ { i } = \\mathrm { A t t e n t i o n } ( Q _ { i } , K _ { i } , V _ { i } ) = \\mathrm { s o f t m a x } \\left( \\frac { Q _ { i } K _ { i } ^ { T } } { \\sqrt { d _ { k } } } \\right) V _ { i }\n$$",
|
| 1718 |
+
"text_format": "latex",
|
| 1719 |
+
"bbox": [
|
| 1720 |
+
323,
|
| 1721 |
+
309,
|
| 1722 |
+
674,
|
| 1723 |
+
352
|
| 1724 |
+
],
|
| 1725 |
+
"page_idx": 12
|
| 1726 |
+
},
|
| 1727 |
+
{
|
| 1728 |
+
"type": "text",
|
| 1729 |
+
"text": "where $Q , K \\in \\mathcal { R } ^ { d _ { k } \\times S }$ , $V \\in \\mathcal { R } ^ { d _ { v } \\times S }$ , $i \\in { 1 , 2 , \\dots , h \\ d _ { k } }$ and $d _ { v }$ are hyper-parameters and all $W \\mathrm { { s } }$ are parameters. The output of all attention heads, $Z _ { i } \\mathrm { s }$ , are concatenated and projected to the same dimension as the input. Finally, the position-wise fully connected $( f _ { \\mathrm { F C } } )$ layer applies two linear transformations to each position with a ReLU activations to the output of the multi-head attention. ",
|
| 1730 |
+
"bbox": [
|
| 1731 |
+
174,
|
| 1732 |
+
354,
|
| 1733 |
+
821,
|
| 1734 |
+
412
|
| 1735 |
+
],
|
| 1736 |
+
"page_idx": 12
|
| 1737 |
+
},
|
| 1738 |
+
{
|
| 1739 |
+
"type": "text",
|
| 1740 |
+
"text": "B DATASET DETAILS ",
|
| 1741 |
+
"text_level": 1,
|
| 1742 |
+
"bbox": [
|
| 1743 |
+
176,
|
| 1744 |
+
431,
|
| 1745 |
+
364,
|
| 1746 |
+
446
|
| 1747 |
+
],
|
| 1748 |
+
"page_idx": 12
|
| 1749 |
+
},
|
| 1750 |
+
{
|
| 1751 |
+
"type": "text",
|
| 1752 |
+
"text": "We generate synthetic datasets for training the spatial planning models for both navigation and manipulation settings. For the navigation setting, we perform experiments with $M \\times M$ maps with two different map sizes, $M \\in \\{ 1 5 , 3 0 \\}$ . We randomly generate $o _ { m i n } = 0$ to $o _ { m a x } = 5$ obstacles in each map, where each obstacle is an rectangle at a random location with each side being a random length from 1 to $M / 2$ . All the rectangular obstacles are rotated in two random orientations. ",
|
| 1753 |
+
"bbox": [
|
| 1754 |
+
174,
|
| 1755 |
+
463,
|
| 1756 |
+
825,
|
| 1757 |
+
534
|
| 1758 |
+
],
|
| 1759 |
+
"page_idx": 12
|
| 1760 |
+
},
|
| 1761 |
+
{
|
| 1762 |
+
"type": "text",
|
| 1763 |
+
"text": "For the manipulation setting, we consider a reacher task using a planar arm with 2 degrees of freedom. We use an operational space of size $P \\times P$ . Each link of the arm is of size $P / 4$ . The arm is centered at the center of the operational space. Let the orientation of two links be denoted by $\\theta _ { 1 }$ and $\\theta _ { 2 }$ . We assume both the links can freely rotate in a plane, $\\theta _ { 1 } , \\theta _ { 2 } \\in [ 0 , 2 \\pi )$ . For each environment, we generate $o _ { m i n } = 0$ to $o _ { m a x } = 5$ circular obstacles centered at a random location $0 . 2 5 P$ to $0 . 7 5 P$ distance away from the center, with a random radius between $0 . 0 5 P$ and $D - 0 . 1 5 P$ where $D$ is the distance of the center of the obstacle from the center of the operational space. We convert each environment to a configuration space map of size $M \\times M$ , where each cell $( i , j )$ denotes whether the arm will collide with an obstacle when $\\theta _ { 1 } = 2 \\pi i / M$ and $\\theta _ { 2 } = 2 \\pi j / M$ . We experiment with two map sizes, $M \\in \\{ 1 8 , 3 6 \\}$ , corresponding to $2 0 ^ { \\circ }$ and $1 0 ^ { \\circ }$ bins for each link. The choice of $P$ does not affect the map as the collision check for each cell in the configuration space is performed in the continuous operational space where all distances are relative to $P$ . ",
|
| 1764 |
+
"bbox": [
|
| 1765 |
+
173,
|
| 1766 |
+
540,
|
| 1767 |
+
826,
|
| 1768 |
+
707
|
| 1769 |
+
],
|
| 1770 |
+
"page_idx": 12
|
| 1771 |
+
},
|
| 1772 |
+
{
|
| 1773 |
+
"type": "text",
|
| 1774 |
+
"text": "C NAVIGATION MAPPER ARCHITECTURE DETAILS ",
|
| 1775 |
+
"text_level": 1,
|
| 1776 |
+
"bbox": [
|
| 1777 |
+
173,
|
| 1778 |
+
727,
|
| 1779 |
+
611,
|
| 1780 |
+
743
|
| 1781 |
+
],
|
| 1782 |
+
"page_idx": 12
|
| 1783 |
+
},
|
| 1784 |
+
{
|
| 1785 |
+
"type": "text",
|
| 1786 |
+
"text": "The Navigation Mapper module predicts a single value between 0 and 1 for each image in $o$ indicating whether the cell in the front of the image is an obstacle or not. The architecture of the Navigation mapper consists of ResNet18 convolutional layers followed by 3 fully-connected layers of size 256, 128 and 1 as shown in Figure 7. Each cell can have upto 4 predictions (from images corresponding to the four neighboring cells facing the current cell), which are aggregated using max pooling to get a single prediction. ",
|
| 1787 |
+
"bbox": [
|
| 1788 |
+
174,
|
| 1789 |
+
758,
|
| 1790 |
+
825,
|
| 1791 |
+
843
|
| 1792 |
+
],
|
| 1793 |
+
"page_idx": 12
|
| 1794 |
+
},
|
| 1795 |
+
{
|
| 1796 |
+
"type": "text",
|
| 1797 |
+
"text": "D HIGHER DIMENSIONAL STATE AND ACTION SPACES ",
|
| 1798 |
+
"text_level": 1,
|
| 1799 |
+
"bbox": [
|
| 1800 |
+
174,
|
| 1801 |
+
862,
|
| 1802 |
+
643,
|
| 1803 |
+
880
|
| 1804 |
+
],
|
| 1805 |
+
"page_idx": 12
|
| 1806 |
+
},
|
| 1807 |
+
{
|
| 1808 |
+
"type": "text",
|
| 1809 |
+
"text": "To test whether SPT maintains performance benefits in higher dimensional state and action spaces, we conducted some experiments for the navigation task. We relaxed the action space from 4 actions to 100 actions by just allowing the agent to take any action in a $1 0 \\mathrm { x } 1 0$ grid around it (and using a low-level controller to go to any cell). The state space used for planning is discretized but the agent moves in a continuous state space in the Habitat simulator. We compute the continuous ground truth distance using Fast Marching Method (instead of Dijkstra) for training with a larger action space, which allows us to accurately compute distance for all locations and not be constrained by axis-aligned actions and distances. ",
|
| 1810 |
+
"bbox": [
|
| 1811 |
+
174,
|
| 1812 |
+
895,
|
| 1813 |
+
823,
|
| 1814 |
+
924
|
| 1815 |
+
],
|
| 1816 |
+
"page_idx": 12
|
| 1817 |
+
},
|
| 1818 |
+
{
|
| 1819 |
+
"type": "image",
|
| 1820 |
+
"img_path": "images/19d0e13b7e28837c7d9cd6c05f4bba84d98b98ea358d21a218421522380e5c8f.jpg",
|
| 1821 |
+
"image_caption": [
|
| 1822 |
+
"Figure 7: Navigation Mapper Architecture. Figure showing the architecture of the Navigation Mapper. "
|
| 1823 |
+
],
|
| 1824 |
+
"image_footnote": [],
|
| 1825 |
+
"bbox": [
|
| 1826 |
+
181,
|
| 1827 |
+
99,
|
| 1828 |
+
823,
|
| 1829 |
+
190
|
| 1830 |
+
],
|
| 1831 |
+
"page_idx": 13
|
| 1832 |
+
},
|
| 1833 |
+
{
|
| 1834 |
+
"type": "text",
|
| 1835 |
+
"text": "",
|
| 1836 |
+
"bbox": [
|
| 1837 |
+
173,
|
| 1838 |
+
239,
|
| 1839 |
+
825,
|
| 1840 |
+
324
|
| 1841 |
+
],
|
| 1842 |
+
"page_idx": 13
|
| 1843 |
+
},
|
| 1844 |
+
{
|
| 1845 |
+
"type": "text",
|
| 1846 |
+
"text": "SPT and the baselines are trained only on synthetic navigation mazes for this experiment. During evaluation, we assume perfect map based on part of the environment seen in the observations so far for planning. If the overall map size at this level of discretization is higher than planning map size, we simply use greedy planning in a window around the agent resulting in an “anytime” variant just like the classical RRT algorithms. This setup results in much finer-grained action space as shown in the demo example here 1. SPT achieve navigation success rate of $7 8 . 0 \\%$ as compared to $4 7 . 2 \\%$ for GPPN and $4 3 . 5 \\%$ for VIN baselines. ",
|
| 1847 |
+
"bbox": [
|
| 1848 |
+
173,
|
| 1849 |
+
330,
|
| 1850 |
+
825,
|
| 1851 |
+
428
|
| 1852 |
+
],
|
| 1853 |
+
"page_idx": 13
|
| 1854 |
+
},
|
| 1855 |
+
{
|
| 1856 |
+
"type": "text",
|
| 1857 |
+
"text": "E ATTENTION VISUALIZATION ",
|
| 1858 |
+
"text_level": 1,
|
| 1859 |
+
"bbox": [
|
| 1860 |
+
176,
|
| 1861 |
+
449,
|
| 1862 |
+
444,
|
| 1863 |
+
464
|
| 1864 |
+
],
|
| 1865 |
+
"page_idx": 13
|
| 1866 |
+
},
|
| 1867 |
+
{
|
| 1868 |
+
"type": "text",
|
| 1869 |
+
"text": "We show the visualization of attention map corresponding to two different locations in Figure 8. Interestingly, we noticed three consistent patterns: a) at least one of the attention head out of eight captures obstacles (left), b) one of the attention heads focuses on goal location (middle), and c) some attention maps focus on nearby obstacles to get accurate planning distance (right). ",
|
| 1870 |
+
"bbox": [
|
| 1871 |
+
174,
|
| 1872 |
+
481,
|
| 1873 |
+
825,
|
| 1874 |
+
536
|
| 1875 |
+
],
|
| 1876 |
+
"page_idx": 13
|
| 1877 |
+
},
|
| 1878 |
+
{
|
| 1879 |
+
"type": "image",
|
| 1880 |
+
"img_path": "images/d99376d7842806d3731bc14c6f47d4900ffbad75f84477bc77eb410022954d95.jpg",
|
| 1881 |
+
"image_caption": [
|
| 1882 |
+
"Figure 8: Attention Visualization. Visualization of the attention heads learned by Spatial Planning Transformers. SPTs learn an attention for each location in the map with respect to every other location. "
|
| 1883 |
+
],
|
| 1884 |
+
"image_footnote": [],
|
| 1885 |
+
"bbox": [
|
| 1886 |
+
174,
|
| 1887 |
+
549,
|
| 1888 |
+
821,
|
| 1889 |
+
638
|
| 1890 |
+
],
|
| 1891 |
+
"page_idx": 13
|
| 1892 |
+
},
|
| 1893 |
+
{
|
| 1894 |
+
"type": "text",
|
| 1895 |
+
"text": "F EXAMPLES ",
|
| 1896 |
+
"text_level": 1,
|
| 1897 |
+
"bbox": [
|
| 1898 |
+
174,
|
| 1899 |
+
709,
|
| 1900 |
+
299,
|
| 1901 |
+
726
|
| 1902 |
+
],
|
| 1903 |
+
"page_idx": 13
|
| 1904 |
+
},
|
| 1905 |
+
{
|
| 1906 |
+
"type": "text",
|
| 1907 |
+
"text": "We show additional examples for navigation task for in-distribution test set (in Figure 9), out-ofdistribution More Obstacles test set (in Figure 10) and Real-World test set (in Figure 11) each with map size $M = 3 0$ . Additional examples for manipulation task are shown for in-distribution test set (in Figure 12) and for out-of-distribution More Obstacles test set (in Figure 13). ",
|
| 1908 |
+
"bbox": [
|
| 1909 |
+
174,
|
| 1910 |
+
742,
|
| 1911 |
+
825,
|
| 1912 |
+
797
|
| 1913 |
+
],
|
| 1914 |
+
"page_idx": 13
|
| 1915 |
+
},
|
| 1916 |
+
{
|
| 1917 |
+
"type": "text",
|
| 1918 |
+
"text": "We also visualize examples for the end-to-end mapping and planning experiments for the manipulation task. We show examples of map and action distance predictions using the SPT model trained with dense and perfect supervision in Figure 14 and with noisy and sparse supervision in Figure 15. ",
|
| 1919 |
+
"bbox": [
|
| 1920 |
+
174,
|
| 1921 |
+
805,
|
| 1922 |
+
823,
|
| 1923 |
+
848
|
| 1924 |
+
],
|
| 1925 |
+
"page_idx": 13
|
| 1926 |
+
},
|
| 1927 |
+
{
|
| 1928 |
+
"type": "image",
|
| 1929 |
+
"img_path": "images/3392b9f02ad8cc7ac59604d5d968af055dfe04495711ae35ea9bc409e86c4083.jpg",
|
| 1930 |
+
"image_caption": [
|
| 1931 |
+
"Figure 9: Navigation in-distribution test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Navigation in-distribution test set for map size $M = 3 0$ . "
|
| 1932 |
+
],
|
| 1933 |
+
"image_footnote": [],
|
| 1934 |
+
"bbox": [
|
| 1935 |
+
176,
|
| 1936 |
+
121,
|
| 1937 |
+
818,
|
| 1938 |
+
434
|
| 1939 |
+
],
|
| 1940 |
+
"page_idx": 14
|
| 1941 |
+
},
|
| 1942 |
+
{
|
| 1943 |
+
"type": "image",
|
| 1944 |
+
"img_path": "images/0d605ee93c04edf7977abfca2e639ca8216bc16aa4670c71cc1b0387f1c91f86.jpg",
|
| 1945 |
+
"image_caption": [
|
| 1946 |
+
"Figure 10: Navigation out-of-distribution More Obstacles test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Navigation out-of-distribution More Obstacles test set for map size $M = 3 0$ . "
|
| 1947 |
+
],
|
| 1948 |
+
"image_footnote": [],
|
| 1949 |
+
"bbox": [
|
| 1950 |
+
176,
|
| 1951 |
+
535,
|
| 1952 |
+
818,
|
| 1953 |
+
847
|
| 1954 |
+
],
|
| 1955 |
+
"page_idx": 14
|
| 1956 |
+
},
|
| 1957 |
+
{
|
| 1958 |
+
"type": "image",
|
| 1959 |
+
"img_path": "images/57ae283d5e288968dd599a88154c51e98c293aa273635277d0b78b27d022b8ca.jpg",
|
| 1960 |
+
"image_caption": [
|
| 1961 |
+
"Figure 11: Navigation out-of-distribution Real-World test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Navigation out-of-distribution Real-World test set for map size $M = 3 0$ . "
|
| 1962 |
+
],
|
| 1963 |
+
"image_footnote": [],
|
| 1964 |
+
"bbox": [
|
| 1965 |
+
179,
|
| 1966 |
+
119,
|
| 1967 |
+
818,
|
| 1968 |
+
429
|
| 1969 |
+
],
|
| 1970 |
+
"page_idx": 15
|
| 1971 |
+
},
|
| 1972 |
+
{
|
| 1973 |
+
"type": "image",
|
| 1974 |
+
"img_path": "images/eeddcd25e6ed6ee42d5a7805e7bcc9ffa930f3deae19ce166e022b27ed97c185.jpg",
|
| 1975 |
+
"image_caption": [
|
| 1976 |
+
"Figure 12: Manipulation in-distribution test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Manipulation indistribution test set for map size $M = 3 6$ . "
|
| 1977 |
+
],
|
| 1978 |
+
"image_footnote": [],
|
| 1979 |
+
"bbox": [
|
| 1980 |
+
179,
|
| 1981 |
+
525,
|
| 1982 |
+
816,
|
| 1983 |
+
849
|
| 1984 |
+
],
|
| 1985 |
+
"page_idx": 15
|
| 1986 |
+
},
|
| 1987 |
+
{
|
| 1988 |
+
"type": "image",
|
| 1989 |
+
"img_path": "images/318a816d502986fd3feb43f7c665aa279f543170e76102bd72f1a9d4ba7f94f0.jpg",
|
| 1990 |
+
"image_caption": [
|
| 1991 |
+
"Figure 13: Manipulation out-of-distribution More Obstacles test set examples. Figure showing 3 examples of the input, the predictions using the proposed SPT model and the baselines, and the ground truth for the Manipulation out-of-distribution More Obstacles test set for map size $M = 3 6$ . "
|
| 1992 |
+
],
|
| 1993 |
+
"image_footnote": [],
|
| 1994 |
+
"bbox": [
|
| 1995 |
+
178,
|
| 1996 |
+
125,
|
| 1997 |
+
816,
|
| 1998 |
+
445
|
| 1999 |
+
],
|
| 2000 |
+
"page_idx": 16
|
| 2001 |
+
},
|
| 2002 |
+
{
|
| 2003 |
+
"type": "image",
|
| 2004 |
+
"img_path": "images/3ff6846fe7476264d5e489b515b5dadd381854d84a3ab942b665d8a0d9eb6716.jpg",
|
| 2005 |
+
"image_caption": [
|
| 2006 |
+
"Figure 14: Dense and Perfect Supervision. Figure showing examples of map and distance predictions using the SPT model trained with dense and perfect action-level supervision. "
|
| 2007 |
+
],
|
| 2008 |
+
"image_footnote": [],
|
| 2009 |
+
"bbox": [
|
| 2010 |
+
176,
|
| 2011 |
+
545,
|
| 2012 |
+
820,
|
| 2013 |
+
859
|
| 2014 |
+
],
|
| 2015 |
+
"page_idx": 16
|
| 2016 |
+
},
|
| 2017 |
+
{
|
| 2018 |
+
"type": "image",
|
| 2019 |
+
"img_path": "images/6d822a799fd1c7d766722bb030ad6c92774a099d9d164c37b4536f31d797b1af.jpg",
|
| 2020 |
+
"image_caption": [
|
| 2021 |
+
"Figure 15: Sparse and Noisy Supervision. Figure showing examples of map and distance predictions using the SPT model trained with sparse and noisy action-level supervision. "
|
| 2022 |
+
],
|
| 2023 |
+
"image_footnote": [],
|
| 2024 |
+
"bbox": [
|
| 2025 |
+
176,
|
| 2026 |
+
330,
|
| 2027 |
+
821,
|
| 2028 |
+
647
|
| 2029 |
+
],
|
| 2030 |
+
"page_idx": 17
|
| 2031 |
+
}
|
| 2032 |
+
]
|
parse/train/n4IMHNb8_f/n4IMHNb8_f_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/n4IMHNb8_f/n4IMHNb8_f_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/yILzFBjR0Y/yILzFBjR0Y.md
ADDED
|
@@ -0,0 +1,309 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# GraphFormers: GNN-nested Transformers for Representation Learning on Textual Graph
|
| 2 |
+
|
| 3 |
+
Junhan Yang♦∗, Zheng Liu♣, Shitao Xiao♠, Chaozhuo Li♣, Defu Lian♦,
|
| 4 |
+
Sanjay Agrawal♥, Amit Singh♥, Guangzhong $\mathbf { S u n ^ { \bullet } }$ , Xing Xie♣
|
| 5 |
+
$\spadesuit$ University of Science and Technology of China, Hefei, China
|
| 6 |
+
♣ Microsoft Research Asia, Beijing, China
|
| 7 |
+
♠ Beijing University of Posts and Telecommunications, Beijing, China
|
| 8 |
+
♥ Microsoft India Development Center, Bengaluru, India
|
| 9 |
+
yangjun2@mail.ustc.edu.cn,
|
| 10 |
+
{zhengliu,cli,siamit,xingx}@microsoft.com,
|
| 11 |
+
stxiao@bupt.edu.cn,
|
| 12 |
+
{liandefu,gzsun}@ustc.edu.cn,
|
| 13 |
+
sanjayiitk0@gmail.com
|
| 14 |
+
|
| 15 |
+
# Abstract
|
| 16 |
+
|
| 17 |
+
The representation learning on textual graph is to generate low-dimensional embeddings for the nodes based on the individual textual features and the neighbourhood information. Recent breakthroughs on pretrained language models and graph neural networks push forward the development of corresponding techniques. The existing works mainly rely on the cascaded model architecture: the textual features of nodes are independently encoded by language models at first; the textual embeddings are aggregated by graph neural networks afterwards. However, the above architecture is limited due to the independent modeling of textual features. In this work, we propose GraphFormers, where layerwise GNN components are nested alongside the transformer blocks of language models. With the proposed architecture, the text encoding and the graph aggregation are fused into an iterative workflow, making each node’s semantic accurately comprehended from the global perspective. In addition, a progressive learning strategy is introduced, where the model is successively trained on manipulated data and original data to reinforce its capability of integrating information on graph. Extensive evaluations are conducted on three large-scale benchmark datasets, where GraphFormers outperform the SOTA baselines with comparable running efficiency. The source code is released at https://github.com/microsoft/GraphFormers .
|
| 18 |
+
|
| 19 |
+
# 1 Introduction
|
| 20 |
+
|
| 21 |
+
The textual graph is a widely existed data format, where each node is annotated with its textual feature. The representation learning on textual graph is to generate low-dimensional node embeddings based on the individual textual features and the information from the neighbourhood. In recent years, the breakthroughs in pretrained language models and graph neural networks contribute to the development of corresponding techniques. Particularly, with pretrained language models, such as BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019a), the underlying semantics of texts can be captured more precisely; at the same time, with graph neural networks, like GraphSage (Hamilton et al., 2017a) and GAT (Velickovi ˇ c et al., 2018), neighbours can be effectively aggregated for more informative node ´ embeddings. It is necessary to combine both techniques for better textual graph representation. As suggested by GraphSage (Hamilton et al., 2017a) and PinSage (Ying et al., 2018), the textual feature can be independently modeled by text encoders and further aggregated by rear-mounted GNNs for the final node embeddings. Such a representation paradigm has been widely adopted by subsequent works on various scenarios (Zhu et al., 2021; Li et al., 2021; Hu et al., 2020; Liu et al., 2019b; Zhou et al., 2019), where GNNs are combined with powerful PLM-based text encoders.
|
| 22 |
+
|
| 23 |
+

|
| 24 |
+
Figure 1: Model architecture comparison (a center node C is connected with two neighbours N1, N2). (A) Cascaded Transformers-GNN: text embeddings are independently generated by language models and aggregated by rear-mounted GNNs. (B) GNN-nested Transformers: the text encoding and graph aggregation are iteratively performed with the layerwise GNNs and Transformers (TRM).
|
| 25 |
+
|
| 26 |
+
The above way of combination is called the “Cascaded Transformers-GNN” architecture (Figure 1 A), as the language models (built upon Transformers) are deployed ahead of the GNN component. With the above architecture, the text encoding and the graph aggregation are performed in two consecutive steps, where there is no information exchange between the nodes when text embeddings are generated. However, the above workflow is defective considering that the linked nodes are correlated, whose underlying semantics can be mutually enhanced. For example, given a node “notes on transformers” and its neighbour “tutorials on machine translation”; by making reference to the whole context, the “transformers” here can be interpreted as a machine learning model, rather than an electric device.
|
| 27 |
+
|
| 28 |
+
Our Work. We propose “GNN-nested Transformers” (GraphFormers), which are highlighted for the fusion of GNNs and language models (Figure 1 B). In GraphFormers, the GNN components are nested alongside the transformer layers (TRM) of language models, where the text encoding and graph aggregation are fused as an iteratively workflow. In each iteration, the linked nodes will exchange information with each other in the layerwise GNN component; thus, each node will be augmented by its neighbourhood information. The transformer component will work on the augmented node features, where increasingly informative node representations can be generated for the next iteration. Compared with the cascaded architecture, GraphFormers achieve more sufficient utilization of the cross-node information on graph, which significantly benefit the representation quality. Given that the layerwise GNN components merely involve simple and effective multi-head attention, GraphFormers preserve comparable running costs as the existing cascaded Transformers-GNN models.
|
| 29 |
+
|
| 30 |
+
On top of the proposed model architecture, we further improve GraphFormers’ representation quality and practicability as follows. Firstly, the training of GraphFormers is likely to be shortcut: in many cases, the center node itself can be “sufficiently informative”, where the training tasks can be accomplished without leveraging the neighbourhood information. As such, GraphFormers may end up with insufficiently trained GNNs. Inspired by recent success of curriculum learning (Bengio et al., 2009), we propose to train the model progressively: the first round of training is performed with manipulated data, where the nodes are randomly polluted; thus, it becomes harder to make prediction merely rely on the center nodes, and the model will be forced to leverage the whole input nodes. The second round of training gets back to the unpolluted data, where the model will be fit into the targeted distribution. Another concern about GraphFormers is that all the linked nodes are mutually dependent in the representation process: once a new node is presented, all the neighbours, regardless of whether they have been processed before, need to be encoded from scratch. As a result, a great deal of unnecessary computations will be incurred. We introduce unidirectional graph attention to alleviate this problem: only the center node is required to make reference to the neighbours, while the neighbour nodes remain independently encoded. By this means, the existing neighbours’ encoding results can be cached and reused, which significantly saves the computation cost.
|
| 31 |
+
|
| 32 |
+
Extensive evaluations are conducted with three million-scale textual graph datasets: DBLP, Wiki and Product, where the representation quality is measured by the link prediction accuracy. According to our experiment results, GraphFormers significantly outperform the SOTA cascaded TransformersGNN baselines with comparable running efficiency.
|
| 33 |
+
|
| 34 |
+
# 2 Related Work
|
| 35 |
+
|
| 36 |
+
The textual graph representation is an important research topic in multiple areas, such as natural language processing, information retrieval and graph learning (Yang et al., 2015; Wang et al., 2016b,a; Yasunaga et al., 2017; Wang et al., 2019a; Xu et al., 2019). To learn high-quality representation for textual graph, techniques on natural language understanding and graph representation need to be jointly leveraged. In recent years, breakthroughs on pretrained language models (PLM) and graph neural networks (GNN) significantly advance the development of corresponding techniques.
|
| 37 |
+
|
| 38 |
+
PLM. The PLMs are proposed to learn universal language models with neural networks trained on large-scale corpus. The early works were based on shallow networks, e.g, word embeddings learned by Skip-Gram (Mikolov et al., 2013) and GloVe (Pennington et al., 2014). In recent years, the backbone networks are being quickly scaled up: from EMLo (Peters et al., 2018), GPT (Radford et al., 2018), to BERT (Devlin et al., 2018), XLNet (Yang et al., 2019), T5 (Raffel et al., 2019), GPT-3 (Brown et al., 2020). The large-scale models, which get fully trained with massive data, demonstrate superior performances on general NLP tasks. One of the most critical usages of PLMs is text representation, where the underlying semantics of texts are captured by low-dimensional embeddings. Such embeddings achieve competitive results on downstream tasks, like text retrieval and classification (Reimers and Gurevych, 2019; Luan et al., 2020; Gao et al., 2021; Su et al., 2021).
|
| 39 |
+
|
| 40 |
+
GNN. Graph neural networks are recognized as powerful tools of modeling graph data (Hamilton et al., 2017b; Zhou et al., 2020). Such methods (e.g., GCN (Kipf and Welling, 2016), GAT (Velickovi ˇ c´ et al., 2018), GraphSage (Hamilton et al., 2017a)) learn effective message passing mechanisms such that information between the nodes can get aggregated for expressive graph representations.
|
| 41 |
+
|
| 42 |
+
Graph neural networks may also incorporate node attributes, like texts; and it’s quite straightforward to leverage GNNs and PLMs for textual graph representation following the “cascaded architecture” suggested by GraphSage (Hamilton et al., 2017a): the node features are independently encoded at first; then, the node embeddings are aggregated via GNNs to generate the final representations. Such a representation paradigm is widely adopted by subsequent works (Zhu et al., 2021; Li et al., 2021; Hu et al., 2020; Liu et al., 2019b; Zhou et al., 2019). However, the above approaches treat the text encoding and graph aggregation as two consecutive steps, where the node-level features are independently processed. Our work is different from these approaches as the text encoding and graph aggregation are fused as an iterative workflow based on the “GNN-nested Transformers”.
|
| 43 |
+
|
| 44 |
+
# 3 GraphFormers
|
| 45 |
+
|
| 46 |
+
In this work, we deal with textual graph data, where each node $x$ is a text. The node $x$ together with its neighbours $N _ { x }$ are denoted as $G _ { x }$ . Our model learns the embedding for node $x$ based on its own textual feature and the information of its neighbourhood $N _ { x }$ . The generated embeddings are expected to capture the relationship between the nodes, i.e., to accurately predict whether two nodes $x _ { q }$ and $x _ { k }$ are connected based on the embedding similarity.
|
| 47 |
+
|
| 48 |
+
# 3.1 GNN-nested Transformers
|
| 49 |
+
|
| 50 |
+
The encoding process of GraphFormers is indicated as follows. The input nodes (the center node and its neighbours) are tokenized into sequences of tokens, with special tokens [CLS] padded in the front, whose states are used for node representation. The input sequences are mapped into the initial embedding sequences $\{ \mathbf { H } _ { g } ^ { 0 } \} _ { G }$ based on the summation of word embeddings and position embeddings. The embedding sequences are encoded by multiple layers of GNN-nested Transformers (shown as Figure 2), where the graph aggregation and text encoding are iteratively performed.
|
| 51 |
+
|
| 52 |
+
• Graph Aggregation in GNN. Each node is enhanced by its neighbourhood information based on the layerwise graph aggregation. For each node in the $l$ -th layer, the first token-level embedding (corresponding to [CLS]) is taken as the node-level embedding: $\mathbf { z } _ { g } ^ { l } \gets \mathbf { H } _ { g } ^ { l } [ 0 ]$ . The node-level embeddings are gathered from all the nodes and passed to the layerwise GNN for graph aggregation. We leverage Multi-Head Attention (MHA) to encode the node-level embeddings $\bar { \mathbf Z } _ { G } ^ { l }$ $( \{ \bar { \mathbf { z } } _ { g } ^ { l } \} _ { G } )$ , similar as GAT (Velickovi ˇ c et al., 2018). For each attention head, the scaled dot-product attention is performed as: ´
|
| 53 |
+
|
| 54 |
+

|
| 55 |
+
Figure 2: GNN-nested Transformers (using the $l$ -th layer for illustration). The graph aggregation is performed in the first place: the node-level embeddings $\{ \mathbf { z } _ { g } ^ { l } \} _ { G }$ are gathered from all the nodes and processed by the GNN component (the leftmost rectangle). The GNN processed node-level embeddings $\{ \hat { \mathbf { z } } _ { g } ^ { l } \} _ { G }$ are dispatched to their original nodes, which forms the graph-augmented tokenlevel embeddings. The graph-augmented token-level embeddings are further encoded by Transformer.
|
| 56 |
+
|
| 57 |
+
$$
|
| 58 |
+
\begin{array} { c } { { \hat { \mathbf { Z } } _ { G } ^ { l } = \mathrm { M H A } ( \mathbf { Z } _ { G } ^ { l } ) ; } } \\ { { \mathrm { M H A } ( \mathbf { Z } _ { G } ^ { l } ) = \mathrm { C o n c a t } ( \mathbf { h e a d } _ { 1 } , . . . , \mathbf { h e a d } _ { h } ) ; } } \\ { { \mathrm { h e a d } _ { j } = \mathrm { s o f t m a x } ( \frac { \mathbf { Q } \mathbf { K } ^ { \mathrm { T } } } { \sqrt { d } } + \mathbf { B } ) \mathbf { V } ; } } \\ { { \mathrm { \mathbf { Q } } = \mathbf { Z } _ { G } ^ { l } \mathbf { W } _ { j } ^ { Q } ; \mathbf { K } = \mathbf { Z } _ { G } ^ { l } \mathbf { W } _ { j } ^ { K } ; \mathbf { V } = \mathbf { Z } _ { G } ^ { l } \mathbf { W } _ { j } ^ { V } ; } } \end{array}
|
| 59 |
+
$$
|
| 60 |
+
|
| 61 |
+
In the above equations, $\mathbf { W } _ { j } ^ { Q } , \mathbf { W } _ { j } ^ { K }$ , and $\mathbf { W } _ { j } ^ { V }$ are the projection matrices of MHA, corresponding to the $j$ -th attention head. A learnable position bias $\mathbf { B }$ is added to the dot-product result; the positions differentiate the relationship between the nodes; i.e., “center-to-center” ( $x$ to $x$ ), “center-to-neighbour” $\scriptstyle { \dot { x } }$ to $N _ { x }$ ), and “neighbour-to-neighbour” ( $N _ { x }$ to $N _ { x }$ ), respectively.
|
| 62 |
+
|
| 63 |
+
Each of the embeddings $\hat { \mathbf { z } } _ { g } ^ { l }$ $( \hat { \mathbf { z } } _ { g } ^ { l } \in \hat { \mathbf { Z } } _ { G } ^ { l } )$ is dispatched to its original node and concatenated $\left( \oplus \right)$ with the token-level embeddings, which gives rise to the graph-augmented token-level embeddings:
|
| 64 |
+
|
| 65 |
+
$$
|
| 66 |
+
\widehat { \mathbf { H } } _ { g } ^ { l } \mathrm { C o n c a t } ( \widehat { \mathbf { z } } _ { g } ^ { l } , \mathbf { H } _ { g } ^ { l } ) .
|
| 67 |
+
$$
|
| 68 |
+
|
| 69 |
+
In this place, the GNN-processed node-level embeddings $\hat { \mathbf { Z } } _ { G } ^ { l }$ can be interpreted as “messagers”, with which the neighbourhood information can be introduced to each of the nodes.
|
| 70 |
+
|
| 71 |
+
• Text Encoding in Transformer. The graph-augmented token-level embeddings $\widehat { \mathbf { H } } _ { g } ^ { l }$ are processed by the transformer component (Vaswani et al., 2017), where the following computations are performed:
|
| 72 |
+
|
| 73 |
+
$$
|
| 74 |
+
\begin{array} { l } { \widehat { \mathbf { H } } _ { g } ^ { l } = \mathrm { L N } ( \mathbf { H } _ { g } ^ { l } + \mathrm { M H A } ^ { a s y } ( \widehat { \mathbf { H } } _ { g } ^ { l } ) ) ; } \\ { \mathbf { H } _ { g } ^ { l + 1 } = \mathrm { L N } ( \widehat { \mathbf { H } } _ { g } ^ { l } + \mathrm { M L P } ( \widehat { \mathbf { H } } _ { g } ^ { l } ) ) . } \end{array}
|
| 75 |
+
$$
|
| 76 |
+
|
| 77 |
+
In the above equations, MLP is the Multi-Layer Projection unit, and LN is the Layer-Norm unit. We use asymmetric Multi-Head Attention $( \mathrm { M H A } ^ { a s y }$ ), where $\mathbf { Q }$ , K, $\mathbf { V }$ are computed as:
|
| 78 |
+
|
| 79 |
+
$$
|
| 80 |
+
\mathbf { Q } = \mathbf { H } _ { g } ^ { l } \mathbf { W } _ { j } ^ { Q } ; ~ \mathbf { K } = \widehat { \mathbf { H } } _ { g } ^ { l } \mathbf { W } _ { j } ^ { K } ; ~ \mathbf { V } = \widehat { \mathbf { H } } _ { g } ^ { l } \mathbf { W } _ { j } ^ { V } .
|
| 81 |
+
$$
|
| 82 |
+
|
| 83 |
+
Therefore, the output sequence $\mathbf { H } _ { g } ^ { l + 1 }$ will be of the same length as the input sequence $\mathbf { H } _ { g } ^ { l }$ . The encoding result will be used as the input token-level embeddings for the next layer. The node-level embedding at the last layer $\mathbf { z } _ { x } ^ { L }$ (i.e., $\bar { \mathbf { H } } _ { g } ^ { L } [ 0 ] )$ will be used as the final node representation.
|
| 84 |
+
|
| 85 |
+
• Workflow. We summarize GraphFormers’ encoding workflow as Algorithm 1. The initial tokenlevel embeddings $\{ \mathbf { H } _ { g } ^ { 0 } \} _ { G }$ are independently encoded by the first Transformer layer $\mathrm { T R M } ^ { 0 }$ . For a $L$ -layer GraphFormers, the graph aggregation and text encoding are iteratively performed for the subsequent $L \mathrm { - } 1$ steps (from 1 to $L - 1 )$ ). In each step, the node-level embeddings $\dot { \mathbf { Z } } _ { G } ^ { \hat { l } }$ are gathered and
|
| 86 |
+
|
| 87 |
+
Input: The input graphs $G$ (consist of the center node $x$ and its neighbours).
|
| 88 |
+
Output: The embedding for the center node $\mathbf { h } _ { x }$ .
|
| 89 |
+
1 begin
|
| 90 |
+
2 for each text $g \in G$ do
|
| 91 |
+
3 $\mathbf { H } _ { g } ^ { 1 } \mathrm { T R M } ^ { 0 } ( \mathbf { H } _ { g } ^ { 0 } ) ;$ ; // Get the initial token-level embeddings
|
| 92 |
+
4 for $l = 1 , . . . , L - 1$ do
|
| 93 |
+
5 $\mathbf { Z } _ { G } ^ { l } \gets \{ \mathbf { z } _ { g } ^ { l } | g \in G \}$ ; // Gather node-level embeddings to GNN
|
| 94 |
+
6 $\hat { \mathbf { Z } } _ { G } ^ { l } \gets \mathrm { G N N } ( \mathbf { Z } _ { G } ^ { l } )$ ; // Graph aggregation in GNN
|
| 95 |
+
7 for each text $g \in G$ do
|
| 96 |
+
8 $\widehat { \mathbf { H } } _ { g } ^ { l } \gets \mathrm { C o n c a t } ( \hat { \mathbf { z } } _ { g } ^ { l } , \mathbf { H } _ { g } ^ { l } )$ ; // Get graph-augmented token-level embeddings
|
| 97 |
+
9 $\mathbf { H } _ { g } ^ { l + 1 } \mathrm { T R M } ^ { l } ( \widehat { \mathbf { H } } _ { g } ^ { l } )$ ; // Text encoding in Transformer
|
| 98 |
+
|
| 99 |
+
Return $\mathbf { h } _ { x } \mathbf { z } _ { x } ^ { L }$ ;
|
| 100 |
+
|
| 101 |
+
processed by the layerwise GNN component. The output node-level embeddings $\hat { \mathbf { Z } } _ { G } ^ { l }$ are dispatched to their original nodes, which generates the graph-augmented token-level embeddings $\widehat { \mathbf { H } } _ { g } ^ { l }$ . The graphaugmented token-level embedding are further processed by the Transformer component. Finally, The node-level embedding (for the center node $x$ ) in the last layer $\mathbf { z } _ { x } ^ { L }$ is taken as our representation result.
|
| 102 |
+
|
| 103 |
+
• Encoding Complexity. Given an input of $M$ nodes, each one has $P$ tokens; the time complexity of each layer’s encoding operation is $O ( \bar { M } ^ { 2 } + M P ^ { 2 } )$ : the graph aggregation takes $O ( M ^ { 2 } )$ , because $M$ node-level embeddings are gathered for multi-head attention; the text encoding takes $\mathrm { \Delta } O ( M P ^ { 2 } )$ , as each of the $M$ node calls for the multi-head attention of $P$ tokens. Compared with Transformers, the GNN’s computation cost is much smaller, mainly because of two reasons: 1) $M ^ { 2 } \ll M P ^ { 2 }$ in general, 2) operations like MLP are not needed in graph aggregation. Therefore, the working efficiency of GraphFormers is close to the cascaded GNN-Transformers as the extra computation cost of layerwise graph aggregation is relatively small. Such a property is also empirically verified in our experiment.
|
| 104 |
+
|
| 105 |
+
# 3.2 Model Simplification: Unidirectional Graph Aggregation
|
| 106 |
+
|
| 107 |
+
One concern about GraphFormers is that the input nodes are mutually dependent on each other during the encoding process. As a result, to generate the embedding for a node, all the related nodes in its neighbourhood need to be encoded from scratch, regardless of whether they have been processed before. Such a property is unfavorable in practice as a great deal of unnecessary computation cost might be incurred (i.e., a node will be repetitively encoded every time it serves as a neighbour node). We leverage a simple but effective simplification, the unidirectional graph aggregation, to address this problem. Particularly, only the center node $x$ is required to make reference to the neighbourhood; while the rest of nodes $N _ { x }$ remain independently encoded all by their own textual features:
|
| 108 |
+
|
| 109 |
+
$$
|
| 110 |
+
\mathbf { H } _ { g } ^ { l + 1 } = \left\{ \mathrm { T R M } ^ { l } ( \widehat { \mathbf { H } _ { x } ^ { l } } ) , g = x ; \right.
|
| 111 |
+
$$
|
| 112 |
+
|
| 113 |
+
Because the encoding of the neighbour nodes is independent of the center node, the intermediate encoding results $\{ \mathbf { z } _ { g } ^ { 1 . . . L } \} _ { N _ { x } }$ can be cached in storage2 and reused in subsequent computations when they are needed. As a result, the nodes can be prevented from being encoded repetitively, which saves a great deal of unnecessary computation cost. We empirically verify that GraphFormers maintain similar performances when the above simplification is introduced.
|
| 114 |
+
|
| 115 |
+
# 3.3 Model Training: Two-Stage Progressive Learning
|
| 116 |
+
|
| 117 |
+
• Training Objective. We take advantage of link prediction as our training task. Given a pair of nodes $q$ and $k$ , the model is learned to predict whether they are connected based on their embedding
|
| 118 |
+
|
| 119 |
+
Table 1: Specifications of the experimental datasets: the number of items, the number of neighbour nodes on average, and the number of training, validation, testing cases.
|
| 120 |
+
|
| 121 |
+
<table><tr><td></td><td>Product</td><td>DBLP</td><td>Wiki</td></tr><tr><td>#Item</td><td>5,643,688</td><td>4,894,081</td><td>4,818.679</td></tr><tr><td>#N</td><td>4.71</td><td>9.31</td><td>8.86</td></tr><tr><td>#Train</td><td>22,146,934</td><td>3,009,506</td><td>7,145,834</td></tr><tr><td>#Valid</td><td>30,000</td><td>60,000</td><td>66,167</td></tr><tr><td>#Test</td><td>306,742</td><td>100,000</td><td>100,000</td></tr></table>
|
| 122 |
+
|
| 123 |
+
similarity. Particularly, the following classification loss is minimized for a positive pair of $q$ and $k$
|
| 124 |
+
|
| 125 |
+
$$
|
| 126 |
+
\mathcal { L } = - \log \frac { \exp ( \langle \mathbf { h } _ { q } , \mathbf { h } _ { k } \rangle ) } { \exp ( \langle \mathbf { h } _ { q } , \mathbf { h } _ { k } \rangle ) + \sum _ { r \in R } \exp ( \langle \mathbf { h } _ { q } , \mathbf { h } _ { r } \rangle ) } .
|
| 127 |
+
$$
|
| 128 |
+
|
| 129 |
+
In the above equation, $\mathbf { h } _ { q }$ and $\mathbf { h } _ { k }$ are the node embeddings; $\langle \cdot \rangle$ denotes the computation of inner product; $R$ stands for the negative samples. In our implementation, we leverage “in-batch negative samples” (Karpukhin et al., 2020; Luan et al., 2020) for the reduction of encoding cost: a positive sample in one training instance will be used as a negative sample in the rest of the training instances within the same mini-batch.
|
| 130 |
+
|
| 131 |
+
• Two-stage Training. In GraphFormers, the information from the center node and neighbour nodes are not treated equally, which may undermine the model’s training effect. Particularly, the center node’s information can be directly utilized, while the neighbourhood information needs to be introduced via three steps: 1) encoded as node-level embeddings, 2) making graph aggregation with the center node, and 3) introduced to center node’s graph augmented token-level embeddings. The message passing pathway can shortcut when the center nodes are “sufficiently informative”, i.e., two nodes are sufficiently similar with each other in terms of their own textual features, such that their connection can be predicted without considering the neighbours. Given the existence of such cases, GraphFormers may end up with well-trained Transformers but insufficiently trained GNNs.
|
| 132 |
+
|
| 133 |
+
To alleviate the above problem, we introduce a warm-up training task, where the link prediction is made based on the polluted input nodes. Particularly, for each input node $g$ , a subset of its tokens $g _ { m }$ will be randomly masked4. As a result, the classification loss becomes:
|
| 134 |
+
|
| 135 |
+
$$
|
| 136 |
+
\mathcal { L } ^ { \prime } = - \log \frac { \exp ( \langle \mathbf { h } _ { \widetilde { q } } , \mathbf { h } _ { \widetilde { k } } \rangle ) } { \exp ( \langle \mathbf { h } _ { \widetilde { q } } , \mathbf { h } _ { \widetilde { k } } \rangle ) + \sum _ { r \in R } \exp ( \langle \mathbf { h } _ { \widetilde { q } } , \mathbf { h } _ { \widetilde { r } } \rangle ) } ,
|
| 137 |
+
$$
|
| 138 |
+
|
| 139 |
+
where $\mathbf { h } _ { \widetilde { q } } , \mathbf { h } _ { \widetilde { k } } .$ , $\mathbf { h } _ { \tilde { r } }$ are the embeddings generated from the polluted nodes. The masked tokens reduce the informativeness of each individual node; therefore, the model is forced to leverage the whole input nodes to make the right prediction.
|
| 140 |
+
|
| 141 |
+
Finally, the model training is organized as a two-stage progressive learning process. In the first stage, the model is trained to minimize $\mathcal { L } ^ { \prime }$ based on the polluted nodes until its convergence, which reinforce the model’s capability of integrating information on graph. In the second stage, the model is continually trained to minimize $\mathcal { L }$ based on the original data until the convergence, which makes the model fit into the target distribution.
|
| 142 |
+
|
| 143 |
+
# 4 Experimental Studies
|
| 144 |
+
|
| 145 |
+
# 4.1 Data and Settings
|
| 146 |
+
|
| 147 |
+
We make use of the following three real-world textual graph datasets for our experimental studies.
|
| 148 |
+
|
| 149 |
+
• $\mathbf { \nabla } \mathbf { D B L P } ^ { 5 }$ , which contains the paper citation graph from DBLP up to 2020-04-09. Two papers are linked if one is cited by the other one. The paper’s title is used as the textual feature.
|
| 150 |
+
|
| 151 |
+
• Wikidata5M6 (Wiki) (Wang et al., 2019b), which contains the entity graph from Wikipedia. The first sentence in each entity’s introduction is taken as its textual feature.
|
| 152 |
+
|
| 153 |
+
Table 2: Overall evaluation (GraphFormers marked in bold, the best baseline underlined). GraphFormers outperforms all baselines, especially the ones based on cascaded Transformers-GNN.
|
| 154 |
+
|
| 155 |
+
<table><tr><td></td><td colspan="3">Product</td><td colspan="3">DBLP</td><td colspan="3">Wiki</td></tr><tr><td>Methods</td><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td></tr><tr><td>PLM</td><td>0.6563</td><td>0.7911</td><td>0.7344</td><td>0.5673</td><td>0.7484</td><td>0.6777</td><td>0.3466</td><td>0.5799</td><td>0.4712</td></tr><tr><td>TNVE</td><td>0.4618</td><td>0.6204</td><td>0.5364</td><td>0.2978</td><td>0.5295</td><td>0.4163</td><td>0.1786</td><td>0.4274</td><td>0.2933</td></tr><tr><td>IFTN</td><td>0.5233</td><td>0.6740</td><td>0.5982</td><td>0.3691</td><td>0.5798</td><td>0.4773</td><td>0.1838</td><td>0.4276</td><td>0.2945</td></tr><tr><td>PLM+GAT</td><td>0.7540</td><td>0.8637</td><td>0.8232</td><td>0.6633</td><td>0.8204</td><td>0.7667</td><td>0.3006</td><td>0.5430</td><td>0.4270</td></tr><tr><td>PLM+Max</td><td>0.7570</td><td>0.8678</td><td>0.8280</td><td>0.6934</td><td>0.8386</td><td>0.7900</td><td>0.3712</td><td>0.6071</td><td>0.5022</td></tr><tr><td>PLM+Mean</td><td>0.7550</td><td>0.8671</td><td>0.8271</td><td>0.6896</td><td>0.8359</td><td>0.7866</td><td>0.3664</td><td>0.6037</td><td>0.4980</td></tr><tr><td>PLM+Att</td><td>0.7513</td><td>0.8652</td><td>0.8246</td><td>0.6910</td><td>0.8366</td><td>0.7875</td><td>0.3709</td><td>0.6067</td><td>0.5018</td></tr><tr><td>GraphFormers</td><td>0.7786</td><td>0.8793</td><td>0.8430</td><td>0.7267</td><td>0.8565</td><td>0.8133</td><td>0.3952</td><td>0.6230</td><td>0.5220</td></tr></table>
|
| 156 |
+
|
| 157 |
+
• Product Graph (Product), an even larger dataset of online products collected by a world-wide search engine. In this dataset, the users’ web browsing behaviors are tracked for the targeted product webpages (e.g., Amazon webpages of Nike shoes). The user’s continuously browsed webpages within a short period of time (e.g., 30 minutes) is called a “session”. The products within a common session are connected in the graph (which is a common way of graph construction in e-commerce scenarios (Ying et al., 2018; Wang et al., 2018)). Each product has its unique textual description, which specifies information like the product name, brand, and saler, etc.
|
| 158 |
+
|
| 159 |
+
The textual features of all the datasets are in English. We make use of uncased WordPiece (Wu et al., 2016) to tokenize the input text. In our experiment, each text is associated with 5 uniformly sampled neighbours (without replacement); for texts with neighbourhood smaller than 5, all the neighbours will be utilized. We summarized the specifications of all the datasets with Table 1. The experiment results are evaluated in terms of link prediction accuracy, i.e., to predict whether a query node and key node are connected given the textual features of themselves and their neighbours. In each testing instance, one query is provided with 300 keys: 1 positive plus 299 randomly sampled negative cases. We leverage three common metrics to measure the prediction accuracy: Precision $@ 1$ , NDCG, and MRR. Without specifications, we will take the unidirectional-simplified GraphFormers trained with the two-stage progressive learning as our default model. More details about the implementations and the training/testing configurations are summarized in an Appendix file. It is submitted together with our source code within the supplementary materials.
|
| 160 |
+
|
| 161 |
+
# 4.2 Baselines
|
| 162 |
+
|
| 163 |
+
We focus on the comparison between GNN-nested Transformers and Cascaded Transformers-GNN. To make sure the difference between both architectures can be truthfully reflected from the evaluation results, GraphFormers and the Cascaded Transformers-GNN baselines are equipped with text encoders and graph aggregators of the same capacities. Particularly, we use the BERT-like PLM as our text encoder, where UniLM-base7 (Bao et al., 2020) is chosen as the network backbone for all related methods; the final layer’s [CLS] token embedding is used for the text embedding.
|
| 164 |
+
|
| 165 |
+
We enumerate the following representative graph aggregators as used in GAT (Velickovi ˇ c et al., 2018), ´ GIN (Xu et al., 2018), GraphSage (Hamilton et al., 2017a). The GAT aggregator, where the node embedding is generated as the weighted sum of all the text embeddings. Each text embedding’s relative importance is calculated as the attention score with the center node. The Pooling-andConcat aggregators, where the center node’s text embedding is concatenated with the neighbours’ pooling result and linearly transformed for the final representation. Depending on the form of pooling function, we have the following options: Max and Mean, where neighbours are aggregated by max-pooling and mean-pooling, respectively; Att, where the neighbours are summed up based on the attention weights with the center node. By comparison, the neighbourhood information may get more emphasized with GAT; while the center node itself tends to be highlighted with Pooling-and-Concat.
|
| 166 |
+
|
| 167 |
+
We consider two more baselines which make use of simplified text encoders (such as CNN) and network embeddings: TNVE (Wang et al., 2019a) and IFTN (Xu et al., 2019). We also include the PLM only baseline, which merely leverages the textual feature of the center node.
|
| 168 |
+
|
| 169 |
+
Table 3: Impact of neighbour size (#N).
|
| 170 |
+
|
| 171 |
+
<table><tr><td rowspan="2">#N</td><td colspan="3">GraphFormers</td><td colspan="3">PLM+Max</td></tr><tr><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td></tr><tr><td>1</td><td>0.6485</td><td>0.8087</td><td>0.7522</td><td>0.6249</td><td>0.7946</td><td>0.7342</td></tr><tr><td>2</td><td>0.6841</td><td>0.8308</td><td>0.7804</td><td>0.6538</td><td>0.8137</td><td>0.7583</td></tr><tr><td>3</td><td>0.6980</td><td>0.8396</td><td>0.7916</td><td>0.6728</td><td>0.8256</td><td>0.7734</td></tr><tr><td>4</td><td>0.7126</td><td>0.8485</td><td>0.8029</td><td>0.6823</td><td>0.8319</td><td>0.7814</td></tr><tr><td>5</td><td>0.7267</td><td>0.8565</td><td>0.8133</td><td>0.6934</td><td>0.8386</td><td>0.7900</td></tr></table>
|
| 172 |
+
|
| 173 |
+
# 4.3 Overall Evaluation
|
| 174 |
+
|
| 175 |
+
The overall evaluation results are reported in Table 2. It’s observed that GraphFormers consistently outperform all the baselines, especially the ones based on the cascaded Transformers-GNN, with notable advantages. Particularly, it achieves $2 . 9 \%$ , $4 . 8 \%$ , $6 . 5 \%$ relative improvements over the most competitive baselines (underlined) on each of the experimental datasets. Such an observation indicates that the relationship between the nodes can be captured more accurately based on the node embeddings generated by GraphFormers, which verifies the effectiveness of our proposed method.
|
| 176 |
+
|
| 177 |
+
We also observe the following underlying factors that may influence the representation quality.
|
| 178 |
+
|
| 179 |
+
Firstly, the effective utilization of neighbourhood information is critical. With the joint consideration of the center node and neighbour nodes, the $\mathrm { P L M + G N N s }$ methods, including GraphFormers and the cascaded Transformers-GNN baselines, significantly outperform the PLM only baseline in most of the time. We further analyze the impact of neighbourhood size as Table 3, with a fraction of neighbour nodes randomly sampled for each center node (using DBLP for illustration). It can be observed that both GraphFormers and $\mathrm { P L M + M a x }$ (the most competitive baseline) achieve higher prediction accuracy than the PLM only method $( \mathrm { P } \ @ 1 . 0 . 5 6 7 3$ , NDCG:0.7484, MRR:0.6777, as reported in Table 2), even with fewer neighbour nodes included. With the increasing number of neighbour nodes, the advantages become gradually enlarged. However, the marginal gain is vanishing, as the relative improvement becomes smaller when more neighbours are included. In all the testing cases, GraphFormers maintain consistent advantages over PLM+Max, which reaffirms the effectiveness of our proposed methods.
|
| 180 |
+
|
| 181 |
+
Secondly, the capacity of the text encoder is crucial for textual graph representation. All the pretrained language model based methods (GraphFormers, Cascaded Transformers-GNN baselines, PLM-only baseline) significantly outperform the baselines with simplified text encoders (TNVE, IFTN).
|
| 182 |
+
|
| 183 |
+
Thirdly, the representation quality is also sensitive to the form of graph aggregator. In Product, the cascaded Transformers-GNN baselines’ performances are quite close to each other. In DBLP, $\mathrm { P L M + }$ (Max, Mean, Att) outperforms PLM+GAT. In Wiki, not only $\mathrm { P L M + }$ (Max, Mean, Att) but also PLM-only baseline outperform PLM $^ +$ GAT. Such phenomenons could be attributed to the type of graph: whether it is homogeneous or heterogeneous. Particularly, both Product and DBLP can be regarded as homogeneous graphs as the nodes are connected based on the same relationships; i.e., co-view relationship in Product, and citation relationship in DBLP. In both homogeneous graphs, the connected nodes may have quite similar semantics (the co-viewed products usually serve similar user intents, and the citation relationships usually indicate similar research topics); thus, the incorporated neighbour nodes will probably provide complementary information for the link prediction between the center nodes. However, Wiki is a heterogeneous graph, where the connections between entities may have highly different semantics. As a result, the incorporation of neighbour nodes may not contribute to the link prediction task, especially when the incorporated neighbours and the prediction target are connected to the center nodes with totally different relationships. Considering that GAT tends to focus more on the neighbourhood, its performance can be vulnerable in such unfavorable situations. These findings suggest that the neighbourhood information should be properly handled in case that the information of the center node is wiped out.
|
| 184 |
+
|
| 185 |
+
Finally, we may conclude different methods’ utility in textual graph representation: simplified text encoders $\prec P L M s \prec C$ ascaded Transformers-GNN $\prec$ GNNs-nested Transformers. Such findings are consistent with our expectation that the precise modeling of individual textual feature and the effective integration of neighbourhood information will jointly contribute to high-quality textual graph representation. GraphFormers enjoy the high expressiveness of PLMs and leverage layerwise nested-GNNs to facilitate graph aggregation, which contributes to both of the above perspectives.
|
| 186 |
+
|
| 187 |
+
Table 4: Ablation Studies (The top ablated methods are marked in bold; $\uparrow ^ { \mathrm { , , } } / \downarrow ^ { \mathrm { , } } \downarrow ^ { \mathrm { , } }$ : the performance is increased/decreased compared with the default setting). “-Progressive”: two-stage progressive learning disabled; “-Simplified”: unidirectional simplification disabled; “-Shared GNNs”: GNNs parameters are not shared across the layers; “-Position”: GNNs learnable position bias disabled.
|
| 188 |
+
|
| 189 |
+
<table><tr><td></td><td colspan="3">Product</td><td colspan="3">DBLP</td><td colspan="3">Wiki</td></tr><tr><td>Methods</td><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td></tr><tr><td>GraphFormers</td><td>0.7786</td><td>0.8793</td><td>0.8430</td><td>0.7267</td><td>0.8565</td><td>0.8133</td><td>0.3952</td><td>0.6230</td><td>0.5220</td></tr><tr><td>PLM+Max</td><td>0.7570</td><td>0.8678</td><td>0.8280</td><td>0.6934</td><td>0.8386</td><td>0.7900</td><td>0.3712</td><td>0.6071</td><td>0.5022</td></tr><tr><td>- Progressive</td><td>0.7688</td><td>0.8751</td><td>0.8373</td><td>0.7096</td><td>0.8468</td><td>0.8007</td><td>0.3834</td><td>0.6155</td><td>0.5127</td></tr><tr><td>- Simplified</td><td>0.7795个</td><td>0.8798 个</td><td>0.8436 个</td><td>0.7225</td><td>0.8542</td><td>0.8102</td><td>0.3923</td><td>0.6209</td><td>0.5195</td></tr><tr><td>- Shared GNNs</td><td>0.7788</td><td>0.8795</td><td>0.8433</td><td>0.7256</td><td>0.8558</td><td>0.8123</td><td>0.3945 ↓</td><td>0.6221↓</td><td>0.5211 ↓</td></tr><tr><td>- Position</td><td>0.7788</td><td>0.8795</td><td>0.8434</td><td>0.7276个</td><td>0.8570个</td><td>0.8139个</td><td>0.3942</td><td>0.6222</td><td>0.5211</td></tr></table>
|
| 190 |
+
|
| 191 |
+
Table 5: Time and memory costs per mini-batch for $\mathrm { P L M + M a x }$ and GraphFormers, with neighbour size increased from 3 to 200. GraphFormers achieve similar efficiency and scalability as $\mathrm { P L M + M a x }$
|
| 192 |
+
|
| 193 |
+
<table><tr><td>#N</td><td>3</td><td>5</td><td>10</td><td>20</td><td>50</td><td>100</td><td>200</td></tr><tr><td>Time: PLM+Max</td><td>60.29 ms</td><td>93.41 ms</td><td>161.40 ms</td><td>295.92 ms</td><td>684.16 ms</td><td>1357.93 ms</td><td>2706.35 ms</td></tr><tr><td>Time: GraphFormers</td><td>63.95 ms</td><td>97.19 ms</td><td>170.16 ms</td><td>306.12 ms</td><td>714.32 ms</td><td>1411.09 ms</td><td>2801.67 ms</td></tr><tr><td>Mem: PLM+Max</td><td>1.33 GiB</td><td>1.39 GiB</td><td>1.55 GiB</td><td>1.82 GiB</td><td>2.67 GiB</td><td>4.09 GiB</td><td>6.92 GiB</td></tr><tr><td>Mem: GraphFormers</td><td>1.33 GiB</td><td>1.39 GiB</td><td>1.55 GiB</td><td>1.83 GiB</td><td>2.70 GiB</td><td>4.28 GiB</td><td>7.33 GiB</td></tr></table>
|
| 194 |
+
|
| 195 |
+
# 4.4 Ablation Studies
|
| 196 |
+
|
| 197 |
+
The ablation studies (as Table 4) are performed to clarify the following issues: 1) the impact of two-stage progressive learning, and 2) the impact of unidirectional-simplified GraphFormers.
|
| 198 |
+
|
| 199 |
+
Firstly, the two-stage progressive learning substantially improves GraphFormers’ representation quality. Without such a training strategy ("-Progressive": training directly on the original data), the model’s performance is decreased by $0 . 9 8 \%$ , $1 . 7 1 \%$ , and $1 . 1 8 \%$ in each of the datasets, respectively.
|
| 200 |
+
|
| 201 |
+
Secondly, the performances between simplified and non-simplified (“-Simplified”) GraphFormers are comparable. In fact, the necessity of graph aggregation is not equivalent for the center node and the neighbour nodes: since the center node is the one for representation, it is much more important to ensure that the center node may extract complementary information from its neighbours. The unidirectional-simplified GraphFormers maintain such a property; thus, there is little impact on the final performances. Such a finding affirms that we may safely leverage the simplified model to save the cost of repetitively encoding the existing neighbours.
|
| 202 |
+
|
| 203 |
+
We make two additional ablation studies. “-Shared GNNs”: the GNNs parameters sharing is disabled, where each layer maintains its own graph aggregator (by default, the layerwise GNN components in GraphFormers share the same set of parameters). “-Position”: the learnable position bias $\mathbf { b }$ in Eq. 1) is disabled in GNNs. We find that model’s performance is little affected from the above changes.
|
| 204 |
+
|
| 205 |
+
# 4.5 Efficiency Analysis
|
| 206 |
+
|
| 207 |
+
We compare the time efficiency between GNN-nested Transformers (GraphFormers) and Cascaded Transformers+GNN (using PLM+Max for comparison). The evaluation is made with a Nvidia P100 GPU. Each mini-batch contains 32 encoding instances; each instance contains one center and #N neighbour nodes; the token length of each node is 16. We report the average time and memory (GPU RAM) costs per mini-batch as Table 5.
|
| 208 |
+
|
| 209 |
+
Firstly, the time and memory costs of both methods grow linearly with the increment of neighbours. (There are overheads of time and memory costs. The time cost overhead may come from CPU processing; while the memory cost overhead is mainly due to the model parameters (Rajbhandari et al., 2020)). We may approximately remove the overheads by deducting the time and memory costs where $\# \mathrm { N } { = } 3$ ). Such a finding is consistent with our theoretical analysis in Section 3.1.
|
| 210 |
+
|
| 211 |
+
Secondly, the overall time and memory costs of GraphFormers are quite close to $\mathrm { P L M + M a x }$ . When the number of neighbour nodes is small, the differences between both methods are almost ignorable. The differences become slightly larger when more neighbour nodes are included, because the layerwise graph aggregations in GraphFormers get increasingly time consuming. However, the differences are still relatively small: merely around $3 . 5 \%$ of the overall running costs when #N is increased to 200 $\because \angle A N = 2 0 0 ^ { \circ }$ is already more than enough for most of the real world scenarios).
|
| 212 |
+
|
| 213 |
+

|
| 214 |
+
Figure 3: Online $\mathrm { A } / \mathrm { B }$ Test: the relative improvements of RPM, CY and CPC against the last version of production system in Bing Search (green: positive; blue: negative). In most of the time, all three performance indicators are significantly improved thanks to the utilization of GraphFormers.
|
| 215 |
+
|
| 216 |
+
Based on the above observations, we may conclude that GraphFormers are more accurate, meanwhile equally efficient and scalable as the conventional cascaded Transformer+GNNs.
|
| 217 |
+
|
| 218 |
+
# 4.6 Online A/B Test on Bing Search
|
| 219 |
+
|
| 220 |
+
GraphFormers has been deployed as one of the major ads retrieval algorithms on Bing Search, and it achieves highly competitive performance against the previous production system (the combination of a wide spectrum of semantic representation algorithms, including large-scale PLMs and cascaded PLMsGNNs). Particularly, the primary objective of Ads service is to maximize the revenue meanwhile increasing the user clicks. Therefore, the following three metrics are taken as the major performance indicators: $\mathrm { R P M } ^ { 8 }$ (revenue per thousand impressions), CY (click yield), and $\mathrm { C P C ^ { 9 } }$ (cost per click) . During our large-scale online $\mathrm { A } / \mathrm { B }$ test, GraphFormers significantly improves the overall RPM, CY, CPC by $1 . 8 7 \%$ , $0 . 9 6 \%$ and $0 . 9 1 \%$ , respectively. A 11-day performance snapshot is demonstrated as Figure 3; it can be observed that in most of the time, all three metrics are significantly improved thanks to the utilization of GraphFormers (the daily performance are measured based on millions of impressions, thus having strong statistic significance).
|
| 221 |
+
|
| 222 |
+
# 5 Conclusion
|
| 223 |
+
|
| 224 |
+
In this paper, we propose a novel model architecture GraphFormers for textual graph representation. By having GNNs nested alongside each transformer layer of the pretrained language model, the underlying semantic of each textual node can be precisely captured and effectively integrated for high-quality textual graph representation. On top of the fundamental architecture, we introduce the two-stage progressive training strategy to further strengthen GraphFormers’ representation quality; we also simplify the model with the unidirectional graph aggregation, which eliminates the unnecessary computation cost. The experimental studies on three large-scale textual graph datasets verify the effectiveness of our proposed methods, where GraphFormers notably outperform the existing cascaded Transformer-GNNs methods with comparable running efficiency and scalability.
|
| 225 |
+
|
| 226 |
+
# 6 Acknowledgement
|
| 227 |
+
|
| 228 |
+
We are grateful to anonymous reviewers for their constructive comments on this work. The work was supported by grants from the National Natural Science Foundation of China (No. 62022077).
|
| 229 |
+
|
| 230 |
+
# References
|
| 231 |
+
|
| 232 |
+
Hangbo Bao, Li Dong, Furu Wei, Wenhui Wang, Nan Yang, Xiaodong Liu, Yu Wang, Jianfeng Gao, Songhao Piao, Ming Zhou, et al. 2020. Unilmv2: Pseudo-masked language models for unified
|
| 233 |
+
|
| 234 |
+
language model pre-training. In International Conference on Machine Learning, pages 642–652.
|
| 235 |
+
PMLR.
|
| 236 |
+
|
| 237 |
+
Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48.
|
| 238 |
+
|
| 239 |
+
Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165.
|
| 240 |
+
|
| 241 |
+
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.
|
| 242 |
+
|
| 243 |
+
Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821.
|
| 244 |
+
|
| 245 |
+
Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017a. Inductive representation learning on large graphs. In Advances in neural information processing systems, pages 1024–1034.
|
| 246 |
+
|
| 247 |
+
William L Hamilton, Rex Ying, and Jure Leskovec. 2017b. Representation learning on graphs: Methods and applications. arXiv preprint arXiv:1709.05584.
|
| 248 |
+
|
| 249 |
+
Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun. 2020. Gpt-gnn: Generative pre-training of graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1857–1867.
|
| 250 |
+
|
| 251 |
+
Vladimir Karpukhin, Barlas Oguz, Sewon Min, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-˘ tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906.
|
| 252 |
+
|
| 253 |
+
Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907.
|
| 254 |
+
|
| 255 |
+
Chaozhuo Li, Bochen Pang, Yuming Liu, Hao Sun, Zheng Liu, Xing Xie, Tianqi Yang, Yanling Cui, Liangjie Zhang, and Qi Zhang. 2021. Adsgnn: Behavior-graph augmented relevance modeling in sponsored search. arXiv preprint arXiv:2104.12080.
|
| 256 |
+
|
| 257 |
+
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019a. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
|
| 258 |
+
|
| 259 |
+
Zhenghao Liu, Chenyan Xiong, Maosong Sun, and Zhiyuan Liu. 2019b. Fine-grained fact verification with kernel graph attention network. arXiv preprint arXiv:1910.09796.
|
| 260 |
+
|
| 261 |
+
Yi Luan, Jacob Eisenstein, Kristina Toutanova, and Michael Collins. 2020. Sparse, dense, and attentional representations for text retrieval. arXiv preprint arXiv:2005.00181.
|
| 262 |
+
|
| 263 |
+
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Distributed representations of words and phrases and their compositionality. arXiv preprint arXiv:1310.4546.
|
| 264 |
+
|
| 265 |
+
Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In 2014 EMNLP, pages 1532–1543.
|
| 266 |
+
|
| 267 |
+
Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. arXiv preprint arXiv:1802.05365.
|
| 268 |
+
|
| 269 |
+
Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training.
|
| 270 |
+
|
| 271 |
+
Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683.
|
| 272 |
+
|
| 273 |
+
Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. 2020. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–16. IEEE.
|
| 274 |
+
|
| 275 |
+
Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bertnetworks. arXiv preprint arXiv:1908.10084.
|
| 276 |
+
|
| 277 |
+
Jianlin Su, Jiarun Cao, Weijie Liu, and Yangyiwen Ou. 2021. Whitening sentence representations for better semantics and faster retrieval. arXiv preprint arXiv:2103.15316.
|
| 278 |
+
|
| 279 |
+
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. arXiv preprint arXiv:1706.03762.
|
| 280 |
+
|
| 281 |
+
Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua ´ Bengio. 2018. Graph attention networks. International Conference on Learning Representations (ICLR).
|
| 282 |
+
|
| 283 |
+
Chenguang Wang, Yangqiu Song, Haoran Li, Ming Zhang, and Jiawei Han. 2016a. Text classification with heterogeneous information network kernels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30.
|
| 284 |
+
|
| 285 |
+
Jizhe Wang, Pipei Huang, Huan Zhao, Zhibo Zhang, Binqiang Zhao, and Dik Lun Lee. 2018. Billionscale commodity embedding for e-commerce recommendation in alibaba. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 839–848.
|
| 286 |
+
|
| 287 |
+
Suhang Wang, Jiliang Tang, Charu Aggarwal, and Huan Liu. 2016b. Linked document embedding for classification. In Proceedings of the 25th ACM international on conference on information and knowledge management, pages 115–124.
|
| 288 |
+
|
| 289 |
+
Wenlin Wang, Chenyang Tao, Zhe Gan, Guoyin Wang, Liqun Chen, Xinyuan Zhang, Ruiyi Zhang, Qian Yang, Ricardo Henao, and Lawrence Carin. 2019a. Improving textual network learning with variational homophilic embeddings. arXiv preprint arXiv:1909.13456.
|
| 290 |
+
|
| 291 |
+
Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhiyuan Liu, Juanzi Li, and Jian Tang. 2019b. Kepler: A unified model for knowledge embedding and pre-trained language representation. arXiv preprint arXiv:1911.06136.
|
| 292 |
+
|
| 293 |
+
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144.
|
| 294 |
+
|
| 295 |
+
Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826.
|
| 296 |
+
|
| 297 |
+
Zenan Xu, Qinliang Su, Xiaojun Quan, and Weijia Zhang. 2019. A deep neural information fusion architecture for textual network embeddings. arXiv preprint arXiv:1908.11057.
|
| 298 |
+
|
| 299 |
+
Cheng Yang, Zhiyuan Liu, Deli Zhao, Maosong Sun, and Edward Y Chang. 2015. Network representation learning with rich text information. In IJCAI, volume 2015, pages 2111–2117.
|
| 300 |
+
|
| 301 |
+
Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint arXiv:1906.08237.
|
| 302 |
+
|
| 303 |
+
Michihiro Yasunaga, Rui Zhang, Kshitijh Meelu, Ayush Pareek, Krishnan Srinivasan, and Dragomir Radev. 2017. Graph-based neural multi-document summarization. arXiv preprint arXiv:1706.06681.
|
| 304 |
+
|
| 305 |
+
Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 974–983.
|
| 306 |
+
|
| 307 |
+
Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2020. Graph neural networks: A review of methods and applications. AI Open, 1:57–81.
|
| 308 |
+
Jie Zhou, Xu Han, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2019. Gear: Graph-based evidence aggregating and reasoning for fact verification. arXiv preprint arXiv:1908.01843.
|
| 309 |
+
Jason Zhu, Yanling Cui, Yuming Liu, Hao Sun, Xue Li, Markus Pelger, Liangjie Zhang, Tianqi Yan, Ruofei Zhang, and Huasha Zhao. 2021. Textgnn: Improving text encoder via graph neural network in sponsored search. arXiv preprint arXiv:2101.06323.
|
parse/train/yILzFBjR0Y/yILzFBjR0Y_content_list.json
ADDED
|
@@ -0,0 +1,1534 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "GraphFormers: GNN-nested Transformers for Representation Learning on Textual Graph ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
218,
|
| 8 |
+
122,
|
| 9 |
+
781,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Junhan Yang♦∗, Zheng Liu♣, Shitao Xiao♠, Chaozhuo Li♣, Defu Lian♦, \nSanjay Agrawal♥, Amit Singh♥, Guangzhong $\\mathbf { S u n ^ { \\bullet } }$ , Xing Xie♣ \n$\\spadesuit$ University of Science and Technology of China, Hefei, China \n♣ Microsoft Research Asia, Beijing, China \n♠ Beijing University of Posts and Telecommunications, Beijing, China \n♥ Microsoft India Development Center, Bengaluru, India \nyangjun2@mail.ustc.edu.cn, \n{zhengliu,cli,siamit,xingx}@microsoft.com, \nstxiao@bupt.edu.cn, \n{liandefu,gzsun}@ustc.edu.cn, \nsanjayiitk0@gmail.com ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
254,
|
| 19 |
+
224,
|
| 20 |
+
748,
|
| 21 |
+
378
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Abstract ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
462,
|
| 31 |
+
414,
|
| 32 |
+
535,
|
| 33 |
+
430
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "The representation learning on textual graph is to generate low-dimensional embeddings for the nodes based on the individual textual features and the neighbourhood information. Recent breakthroughs on pretrained language models and graph neural networks push forward the development of corresponding techniques. The existing works mainly rely on the cascaded model architecture: the textual features of nodes are independently encoded by language models at first; the textual embeddings are aggregated by graph neural networks afterwards. However, the above architecture is limited due to the independent modeling of textual features. In this work, we propose GraphFormers, where layerwise GNN components are nested alongside the transformer blocks of language models. With the proposed architecture, the text encoding and the graph aggregation are fused into an iterative workflow, making each node’s semantic accurately comprehended from the global perspective. In addition, a progressive learning strategy is introduced, where the model is successively trained on manipulated data and original data to reinforce its capability of integrating information on graph. Extensive evaluations are conducted on three large-scale benchmark datasets, where GraphFormers outperform the SOTA baselines with comparable running efficiency. The source code is released at https://github.com/microsoft/GraphFormers . ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
233,
|
| 42 |
+
445,
|
| 43 |
+
766,
|
| 44 |
+
694
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "1 Introduction ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
174,
|
| 54 |
+
719,
|
| 55 |
+
310,
|
| 56 |
+
738
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "The textual graph is a widely existed data format, where each node is annotated with its textual feature. The representation learning on textual graph is to generate low-dimensional node embeddings based on the individual textual features and the information from the neighbourhood. In recent years, the breakthroughs in pretrained language models and graph neural networks contribute to the development of corresponding techniques. Particularly, with pretrained language models, such as BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019a), the underlying semantics of texts can be captured more precisely; at the same time, with graph neural networks, like GraphSage (Hamilton et al., 2017a) and GAT (Velickovi ˇ c et al., 2018), neighbours can be effectively aggregated for more informative node ´ embeddings. It is necessary to combine both techniques for better textual graph representation. As suggested by GraphSage (Hamilton et al., 2017a) and PinSage (Ying et al., 2018), the textual feature can be independently modeled by text encoders and further aggregated by rear-mounted GNNs for the final node embeddings. Such a representation paradigm has been widely adopted by subsequent works on various scenarios (Zhu et al., 2021; Li et al., 2021; Hu et al., 2020; Liu et al., 2019b; Zhou et al., 2019), where GNNs are combined with powerful PLM-based text encoders. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
752,
|
| 66 |
+
825,
|
| 67 |
+
877
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "image",
|
| 73 |
+
"img_path": "images/f5a6dedd28f3b25ea18fad08789f31513edfc74d50c8d6fd1c909c50a15b27a6.jpg",
|
| 74 |
+
"image_caption": [
|
| 75 |
+
"Figure 1: Model architecture comparison (a center node C is connected with two neighbours N1, N2). (A) Cascaded Transformers-GNN: text embeddings are independently generated by language models and aggregated by rear-mounted GNNs. (B) GNN-nested Transformers: the text encoding and graph aggregation are iteratively performed with the layerwise GNNs and Transformers (TRM). "
|
| 76 |
+
],
|
| 77 |
+
"image_footnote": [],
|
| 78 |
+
"bbox": [
|
| 79 |
+
174,
|
| 80 |
+
88,
|
| 81 |
+
821,
|
| 82 |
+
253
|
| 83 |
+
],
|
| 84 |
+
"page_idx": 1
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"type": "text",
|
| 88 |
+
"text": "",
|
| 89 |
+
"bbox": [
|
| 90 |
+
174,
|
| 91 |
+
334,
|
| 92 |
+
825,
|
| 93 |
+
404
|
| 94 |
+
],
|
| 95 |
+
"page_idx": 1
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"type": "text",
|
| 99 |
+
"text": "The above way of combination is called the “Cascaded Transformers-GNN” architecture (Figure 1 A), as the language models (built upon Transformers) are deployed ahead of the GNN component. With the above architecture, the text encoding and the graph aggregation are performed in two consecutive steps, where there is no information exchange between the nodes when text embeddings are generated. However, the above workflow is defective considering that the linked nodes are correlated, whose underlying semantics can be mutually enhanced. For example, given a node “notes on transformers” and its neighbour “tutorials on machine translation”; by making reference to the whole context, the “transformers” here can be interpreted as a machine learning model, rather than an electric device. ",
|
| 100 |
+
"bbox": [
|
| 101 |
+
174,
|
| 102 |
+
410,
|
| 103 |
+
826,
|
| 104 |
+
522
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "Our Work. We propose “GNN-nested Transformers” (GraphFormers), which are highlighted for the fusion of GNNs and language models (Figure 1 B). In GraphFormers, the GNN components are nested alongside the transformer layers (TRM) of language models, where the text encoding and graph aggregation are fused as an iteratively workflow. In each iteration, the linked nodes will exchange information with each other in the layerwise GNN component; thus, each node will be augmented by its neighbourhood information. The transformer component will work on the augmented node features, where increasingly informative node representations can be generated for the next iteration. Compared with the cascaded architecture, GraphFormers achieve more sufficient utilization of the cross-node information on graph, which significantly benefit the representation quality. Given that the layerwise GNN components merely involve simple and effective multi-head attention, GraphFormers preserve comparable running costs as the existing cascaded Transformers-GNN models. ",
|
| 111 |
+
"bbox": [
|
| 112 |
+
173,
|
| 113 |
+
527,
|
| 114 |
+
825,
|
| 115 |
+
680
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "On top of the proposed model architecture, we further improve GraphFormers’ representation quality and practicability as follows. Firstly, the training of GraphFormers is likely to be shortcut: in many cases, the center node itself can be “sufficiently informative”, where the training tasks can be accomplished without leveraging the neighbourhood information. As such, GraphFormers may end up with insufficiently trained GNNs. Inspired by recent success of curriculum learning (Bengio et al., 2009), we propose to train the model progressively: the first round of training is performed with manipulated data, where the nodes are randomly polluted; thus, it becomes harder to make prediction merely rely on the center nodes, and the model will be forced to leverage the whole input nodes. The second round of training gets back to the unpolluted data, where the model will be fit into the targeted distribution. Another concern about GraphFormers is that all the linked nodes are mutually dependent in the representation process: once a new node is presented, all the neighbours, regardless of whether they have been processed before, need to be encoded from scratch. As a result, a great deal of unnecessary computations will be incurred. We introduce unidirectional graph attention to alleviate this problem: only the center node is required to make reference to the neighbours, while the neighbour nodes remain independently encoded. By this means, the existing neighbours’ encoding results can be cached and reused, which significantly saves the computation cost. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
686,
|
| 125 |
+
825,
|
| 126 |
+
906
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "Extensive evaluations are conducted with three million-scale textual graph datasets: DBLP, Wiki and Product, where the representation quality is measured by the link prediction accuracy. According to our experiment results, GraphFormers significantly outperform the SOTA cascaded TransformersGNN baselines with comparable running efficiency. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
90,
|
| 136 |
+
825,
|
| 137 |
+
147
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 2
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "2 Related Work ",
|
| 144 |
+
"text_level": 1,
|
| 145 |
+
"bbox": [
|
| 146 |
+
174,
|
| 147 |
+
165,
|
| 148 |
+
321,
|
| 149 |
+
183
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 2
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "text",
|
| 155 |
+
"text": "The textual graph representation is an important research topic in multiple areas, such as natural language processing, information retrieval and graph learning (Yang et al., 2015; Wang et al., 2016b,a; Yasunaga et al., 2017; Wang et al., 2019a; Xu et al., 2019). To learn high-quality representation for textual graph, techniques on natural language understanding and graph representation need to be jointly leveraged. In recent years, breakthroughs on pretrained language models (PLM) and graph neural networks (GNN) significantly advance the development of corresponding techniques. ",
|
| 156 |
+
"bbox": [
|
| 157 |
+
173,
|
| 158 |
+
196,
|
| 159 |
+
825,
|
| 160 |
+
281
|
| 161 |
+
],
|
| 162 |
+
"page_idx": 2
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"type": "text",
|
| 166 |
+
"text": "PLM. The PLMs are proposed to learn universal language models with neural networks trained on large-scale corpus. The early works were based on shallow networks, e.g, word embeddings learned by Skip-Gram (Mikolov et al., 2013) and GloVe (Pennington et al., 2014). In recent years, the backbone networks are being quickly scaled up: from EMLo (Peters et al., 2018), GPT (Radford et al., 2018), to BERT (Devlin et al., 2018), XLNet (Yang et al., 2019), T5 (Raffel et al., 2019), GPT-3 (Brown et al., 2020). The large-scale models, which get fully trained with massive data, demonstrate superior performances on general NLP tasks. One of the most critical usages of PLMs is text representation, where the underlying semantics of texts are captured by low-dimensional embeddings. Such embeddings achieve competitive results on downstream tasks, like text retrieval and classification (Reimers and Gurevych, 2019; Luan et al., 2020; Gao et al., 2021; Su et al., 2021). ",
|
| 167 |
+
"bbox": [
|
| 168 |
+
173,
|
| 169 |
+
286,
|
| 170 |
+
825,
|
| 171 |
+
426
|
| 172 |
+
],
|
| 173 |
+
"page_idx": 2
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"type": "text",
|
| 177 |
+
"text": "GNN. Graph neural networks are recognized as powerful tools of modeling graph data (Hamilton et al., 2017b; Zhou et al., 2020). Such methods (e.g., GCN (Kipf and Welling, 2016), GAT (Velickovi ˇ c´ et al., 2018), GraphSage (Hamilton et al., 2017a)) learn effective message passing mechanisms such that information between the nodes can get aggregated for expressive graph representations. ",
|
| 178 |
+
"bbox": [
|
| 179 |
+
174,
|
| 180 |
+
431,
|
| 181 |
+
825,
|
| 182 |
+
487
|
| 183 |
+
],
|
| 184 |
+
"page_idx": 2
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"type": "text",
|
| 188 |
+
"text": "Graph neural networks may also incorporate node attributes, like texts; and it’s quite straightforward to leverage GNNs and PLMs for textual graph representation following the “cascaded architecture” suggested by GraphSage (Hamilton et al., 2017a): the node features are independently encoded at first; then, the node embeddings are aggregated via GNNs to generate the final representations. Such a representation paradigm is widely adopted by subsequent works (Zhu et al., 2021; Li et al., 2021; Hu et al., 2020; Liu et al., 2019b; Zhou et al., 2019). However, the above approaches treat the text encoding and graph aggregation as two consecutive steps, where the node-level features are independently processed. Our work is different from these approaches as the text encoding and graph aggregation are fused as an iterative workflow based on the “GNN-nested Transformers”. ",
|
| 189 |
+
"bbox": [
|
| 190 |
+
174,
|
| 191 |
+
493,
|
| 192 |
+
826,
|
| 193 |
+
617
|
| 194 |
+
],
|
| 195 |
+
"page_idx": 2
|
| 196 |
+
},
|
| 197 |
+
{
|
| 198 |
+
"type": "text",
|
| 199 |
+
"text": "3 GraphFormers ",
|
| 200 |
+
"text_level": 1,
|
| 201 |
+
"bbox": [
|
| 202 |
+
176,
|
| 203 |
+
637,
|
| 204 |
+
331,
|
| 205 |
+
655
|
| 206 |
+
],
|
| 207 |
+
"page_idx": 2
|
| 208 |
+
},
|
| 209 |
+
{
|
| 210 |
+
"type": "text",
|
| 211 |
+
"text": "In this work, we deal with textual graph data, where each node $x$ is a text. The node $x$ together with its neighbours $N _ { x }$ are denoted as $G _ { x }$ . Our model learns the embedding for node $x$ based on its own textual feature and the information of its neighbourhood $N _ { x }$ . The generated embeddings are expected to capture the relationship between the nodes, i.e., to accurately predict whether two nodes $x _ { q }$ and $x _ { k }$ are connected based on the embedding similarity. ",
|
| 212 |
+
"bbox": [
|
| 213 |
+
174,
|
| 214 |
+
669,
|
| 215 |
+
825,
|
| 216 |
+
738
|
| 217 |
+
],
|
| 218 |
+
"page_idx": 2
|
| 219 |
+
},
|
| 220 |
+
{
|
| 221 |
+
"type": "text",
|
| 222 |
+
"text": "3.1 GNN-nested Transformers ",
|
| 223 |
+
"text_level": 1,
|
| 224 |
+
"bbox": [
|
| 225 |
+
174,
|
| 226 |
+
753,
|
| 227 |
+
398,
|
| 228 |
+
768
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 2
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "text",
|
| 234 |
+
"text": "The encoding process of GraphFormers is indicated as follows. The input nodes (the center node and its neighbours) are tokenized into sequences of tokens, with special tokens [CLS] padded in the front, whose states are used for node representation. The input sequences are mapped into the initial embedding sequences $\\{ \\mathbf { H } _ { g } ^ { 0 } \\} _ { G }$ based on the summation of word embeddings and position embeddings. The embedding sequences are encoded by multiple layers of GNN-nested Transformers (shown as Figure 2), where the graph aggregation and text encoding are iteratively performed. ",
|
| 235 |
+
"bbox": [
|
| 236 |
+
174,
|
| 237 |
+
779,
|
| 238 |
+
825,
|
| 239 |
+
863
|
| 240 |
+
],
|
| 241 |
+
"page_idx": 2
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"type": "text",
|
| 245 |
+
"text": "• Graph Aggregation in GNN. Each node is enhanced by its neighbourhood information based on the layerwise graph aggregation. For each node in the $l$ -th layer, the first token-level embedding (corresponding to [CLS]) is taken as the node-level embedding: $\\mathbf { z } _ { g } ^ { l } \\gets \\mathbf { H } _ { g } ^ { l } [ 0 ]$ . The node-level embeddings are gathered from all the nodes and passed to the layerwise GNN for graph aggregation. We leverage Multi-Head Attention (MHA) to encode the node-level embeddings $\\bar { \\mathbf Z } _ { G } ^ { l }$ $( \\{ \\bar { \\mathbf { z } } _ { g } ^ { l } \\} _ { G } )$ , similar as GAT (Velickovi ˇ c et al., 2018). For each attention head, the scaled dot-product attention is performed as: ´ ",
|
| 246 |
+
"bbox": [
|
| 247 |
+
174,
|
| 248 |
+
869,
|
| 249 |
+
825,
|
| 250 |
+
912
|
| 251 |
+
],
|
| 252 |
+
"page_idx": 2
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"type": "image",
|
| 256 |
+
"img_path": "images/59392899ccd4deb6b982ce5d2f5abba3ec1f35303b744990a5fa538fc2f0ebc5.jpg",
|
| 257 |
+
"image_caption": [
|
| 258 |
+
"Figure 2: GNN-nested Transformers (using the $l$ -th layer for illustration). The graph aggregation is performed in the first place: the node-level embeddings $\\{ \\mathbf { z } _ { g } ^ { l } \\} _ { G }$ are gathered from all the nodes and processed by the GNN component (the leftmost rectangle). The GNN processed node-level embeddings $\\{ \\hat { \\mathbf { z } } _ { g } ^ { l } \\} _ { G }$ are dispatched to their original nodes, which forms the graph-augmented tokenlevel embeddings. The graph-augmented token-level embeddings are further encoded by Transformer. "
|
| 259 |
+
],
|
| 260 |
+
"image_footnote": [],
|
| 261 |
+
"bbox": [
|
| 262 |
+
178,
|
| 263 |
+
88,
|
| 264 |
+
830,
|
| 265 |
+
231
|
| 266 |
+
],
|
| 267 |
+
"page_idx": 3
|
| 268 |
+
},
|
| 269 |
+
{
|
| 270 |
+
"type": "text",
|
| 271 |
+
"text": "",
|
| 272 |
+
"bbox": [
|
| 273 |
+
173,
|
| 274 |
+
327,
|
| 275 |
+
826,
|
| 276 |
+
371
|
| 277 |
+
],
|
| 278 |
+
"page_idx": 3
|
| 279 |
+
},
|
| 280 |
+
{
|
| 281 |
+
"type": "equation",
|
| 282 |
+
"img_path": "images/8bf5ddb23aba2488ba1563e5fe9ff36712dd269cbc1f6d969b28bb60e5b05b09.jpg",
|
| 283 |
+
"text": "$$\n\\begin{array} { c } { { \\hat { \\mathbf { Z } } _ { G } ^ { l } = \\mathrm { M H A } ( \\mathbf { Z } _ { G } ^ { l } ) ; } } \\\\ { { \\mathrm { M H A } ( \\mathbf { Z } _ { G } ^ { l } ) = \\mathrm { C o n c a t } ( \\mathbf { h e a d } _ { 1 } , . . . , \\mathbf { h e a d } _ { h } ) ; } } \\\\ { { \\mathrm { h e a d } _ { j } = \\mathrm { s o f t m a x } ( \\frac { \\mathbf { Q } \\mathbf { K } ^ { \\mathrm { T } } } { \\sqrt { d } } + \\mathbf { B } ) \\mathbf { V } ; } } \\\\ { { \\mathrm { \\mathbf { Q } } = \\mathbf { Z } _ { G } ^ { l } \\mathbf { W } _ { j } ^ { Q } ; \\mathbf { K } = \\mathbf { Z } _ { G } ^ { l } \\mathbf { W } _ { j } ^ { K } ; \\mathbf { V } = \\mathbf { Z } _ { G } ^ { l } \\mathbf { W } _ { j } ^ { V } ; } } \\end{array}\n$$",
|
| 284 |
+
"text_format": "latex",
|
| 285 |
+
"bbox": [
|
| 286 |
+
348,
|
| 287 |
+
376,
|
| 288 |
+
647,
|
| 289 |
+
477
|
| 290 |
+
],
|
| 291 |
+
"page_idx": 3
|
| 292 |
+
},
|
| 293 |
+
{
|
| 294 |
+
"type": "text",
|
| 295 |
+
"text": "In the above equations, $\\mathbf { W } _ { j } ^ { Q } , \\mathbf { W } _ { j } ^ { K }$ , and $\\mathbf { W } _ { j } ^ { V }$ are the projection matrices of MHA, corresponding to the $j$ -th attention head. A learnable position bias $\\mathbf { B }$ is added to the dot-product result; the positions differentiate the relationship between the nodes; i.e., “center-to-center” ( $x$ to $x$ ), “center-to-neighbour” $\\scriptstyle { \\dot { x } }$ to $N _ { x }$ ), and “neighbour-to-neighbour” ( $N _ { x }$ to $N _ { x }$ ), respectively. ",
|
| 296 |
+
"bbox": [
|
| 297 |
+
176,
|
| 298 |
+
484,
|
| 299 |
+
826,
|
| 300 |
+
544
|
| 301 |
+
],
|
| 302 |
+
"page_idx": 3
|
| 303 |
+
},
|
| 304 |
+
{
|
| 305 |
+
"type": "text",
|
| 306 |
+
"text": "Each of the embeddings $\\hat { \\mathbf { z } } _ { g } ^ { l }$ $( \\hat { \\mathbf { z } } _ { g } ^ { l } \\in \\hat { \\mathbf { Z } } _ { G } ^ { l } )$ is dispatched to its original node and concatenated $\\left( \\oplus \\right)$ with the token-level embeddings, which gives rise to the graph-augmented token-level embeddings: ",
|
| 307 |
+
"bbox": [
|
| 308 |
+
173,
|
| 309 |
+
549,
|
| 310 |
+
825,
|
| 311 |
+
580
|
| 312 |
+
],
|
| 313 |
+
"page_idx": 3
|
| 314 |
+
},
|
| 315 |
+
{
|
| 316 |
+
"type": "equation",
|
| 317 |
+
"img_path": "images/1346106899da82786ab1bd070b832e949b2cb25226f015bc90367e5c945c7083.jpg",
|
| 318 |
+
"text": "$$\n\\widehat { \\mathbf { H } } _ { g } ^ { l } \\mathrm { C o n c a t } ( \\widehat { \\mathbf { z } } _ { g } ^ { l } , \\mathbf { H } _ { g } ^ { l } ) .\n$$",
|
| 319 |
+
"text_format": "latex",
|
| 320 |
+
"bbox": [
|
| 321 |
+
415,
|
| 322 |
+
587,
|
| 323 |
+
581,
|
| 324 |
+
608
|
| 325 |
+
],
|
| 326 |
+
"page_idx": 3
|
| 327 |
+
},
|
| 328 |
+
{
|
| 329 |
+
"type": "text",
|
| 330 |
+
"text": "In this place, the GNN-processed node-level embeddings $\\hat { \\mathbf { Z } } _ { G } ^ { l }$ can be interpreted as “messagers”, with which the neighbourhood information can be introduced to each of the nodes. ",
|
| 331 |
+
"bbox": [
|
| 332 |
+
171,
|
| 333 |
+
617,
|
| 334 |
+
825,
|
| 335 |
+
647
|
| 336 |
+
],
|
| 337 |
+
"page_idx": 3
|
| 338 |
+
},
|
| 339 |
+
{
|
| 340 |
+
"type": "text",
|
| 341 |
+
"text": "• Text Encoding in Transformer. The graph-augmented token-level embeddings $\\widehat { \\mathbf { H } } _ { g } ^ { l }$ are processed by the transformer component (Vaswani et al., 2017), where the following computations are performed: ",
|
| 342 |
+
"bbox": [
|
| 343 |
+
169,
|
| 344 |
+
654,
|
| 345 |
+
823,
|
| 346 |
+
684
|
| 347 |
+
],
|
| 348 |
+
"page_idx": 3
|
| 349 |
+
},
|
| 350 |
+
{
|
| 351 |
+
"type": "equation",
|
| 352 |
+
"img_path": "images/c44bcdc2d433b174b33cbdeeeb10a185ef1e8689dcac0497cbae02ba8efed6aa.jpg",
|
| 353 |
+
"text": "$$\n\\begin{array} { l } { \\widehat { \\mathbf { H } } _ { g } ^ { l } = \\mathrm { L N } ( \\mathbf { H } _ { g } ^ { l } + \\mathrm { M H A } ^ { a s y } ( \\widehat { \\mathbf { H } } _ { g } ^ { l } ) ) ; } \\\\ { \\mathbf { H } _ { g } ^ { l + 1 } = \\mathrm { L N } ( \\widehat { \\mathbf { H } } _ { g } ^ { l } + \\mathrm { M L P } ( \\widehat { \\mathbf { H } } _ { g } ^ { l } ) ) . } \\end{array}\n$$",
|
| 354 |
+
"text_format": "latex",
|
| 355 |
+
"bbox": [
|
| 356 |
+
385,
|
| 357 |
+
693,
|
| 358 |
+
611,
|
| 359 |
+
738
|
| 360 |
+
],
|
| 361 |
+
"page_idx": 3
|
| 362 |
+
},
|
| 363 |
+
{
|
| 364 |
+
"type": "text",
|
| 365 |
+
"text": "In the above equations, MLP is the Multi-Layer Projection unit, and LN is the Layer-Norm unit. We use asymmetric Multi-Head Attention $( \\mathrm { M H A } ^ { a s y }$ ), where $\\mathbf { Q }$ , K, $\\mathbf { V }$ are computed as: ",
|
| 366 |
+
"bbox": [
|
| 367 |
+
173,
|
| 368 |
+
739,
|
| 369 |
+
823,
|
| 370 |
+
768
|
| 371 |
+
],
|
| 372 |
+
"page_idx": 3
|
| 373 |
+
},
|
| 374 |
+
{
|
| 375 |
+
"type": "equation",
|
| 376 |
+
"img_path": "images/60d836c623aa4ffaa803c8b5de1e2a4a999f779436e7deff80cfefe9472c1b43.jpg",
|
| 377 |
+
"text": "$$\n\\mathbf { Q } = \\mathbf { H } _ { g } ^ { l } \\mathbf { W } _ { j } ^ { Q } ; ~ \\mathbf { K } = \\widehat { \\mathbf { H } } _ { g } ^ { l } \\mathbf { W } _ { j } ^ { K } ; ~ \\mathbf { V } = \\widehat { \\mathbf { H } } _ { g } ^ { l } \\mathbf { W } _ { j } ^ { V } .\n$$",
|
| 378 |
+
"text_format": "latex",
|
| 379 |
+
"bbox": [
|
| 380 |
+
348,
|
| 381 |
+
775,
|
| 382 |
+
648,
|
| 383 |
+
797
|
| 384 |
+
],
|
| 385 |
+
"page_idx": 3
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"type": "text",
|
| 389 |
+
"text": "Therefore, the output sequence $\\mathbf { H } _ { g } ^ { l + 1 }$ will be of the same length as the input sequence $\\mathbf { H } _ { g } ^ { l }$ . The encoding result will be used as the input token-level embeddings for the next layer. The node-level embedding at the last layer $\\mathbf { z } _ { x } ^ { L }$ (i.e., $\\bar { \\mathbf { H } } _ { g } ^ { L } [ 0 ] )$ will be used as the final node representation. ",
|
| 390 |
+
"bbox": [
|
| 391 |
+
173,
|
| 392 |
+
804,
|
| 393 |
+
825,
|
| 394 |
+
849
|
| 395 |
+
],
|
| 396 |
+
"page_idx": 3
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"type": "text",
|
| 400 |
+
"text": "• Workflow. We summarize GraphFormers’ encoding workflow as Algorithm 1. The initial tokenlevel embeddings $\\{ \\mathbf { H } _ { g } ^ { 0 } \\} _ { G }$ are independently encoded by the first Transformer layer $\\mathrm { T R M } ^ { 0 }$ . For a $L$ -layer GraphFormers, the graph aggregation and text encoding are iteratively performed for the subsequent $L \\mathrm { - } 1$ steps (from 1 to $L - 1 )$ ). In each step, the node-level embeddings $\\dot { \\mathbf { Z } } _ { G } ^ { \\hat { l } }$ are gathered and ",
|
| 401 |
+
"bbox": [
|
| 402 |
+
174,
|
| 403 |
+
853,
|
| 404 |
+
825,
|
| 405 |
+
912
|
| 406 |
+
],
|
| 407 |
+
"page_idx": 3
|
| 408 |
+
},
|
| 409 |
+
{
|
| 410 |
+
"type": "text",
|
| 411 |
+
"text": "Input: The input graphs $G$ (consist of the center node $x$ and its neighbours). \nOutput: The embedding for the center node $\\mathbf { h } _ { x }$ . \n1 begin \n2 for each text $g \\in G$ do \n3 $\\mathbf { H } _ { g } ^ { 1 } \\mathrm { T R M } ^ { 0 } ( \\mathbf { H } _ { g } ^ { 0 } ) ;$ ; // Get the initial token-level embeddings \n4 for $l = 1 , . . . , L - 1$ do \n5 $\\mathbf { Z } _ { G } ^ { l } \\gets \\{ \\mathbf { z } _ { g } ^ { l } | g \\in G \\}$ ; // Gather node-level embeddings to GNN \n6 $\\hat { \\mathbf { Z } } _ { G } ^ { l } \\gets \\mathrm { G N N } ( \\mathbf { Z } _ { G } ^ { l } )$ ; // Graph aggregation in GNN \n7 for each text $g \\in G$ do \n8 $\\widehat { \\mathbf { H } } _ { g } ^ { l } \\gets \\mathrm { C o n c a t } ( \\hat { \\mathbf { z } } _ { g } ^ { l } , \\mathbf { H } _ { g } ^ { l } )$ ; // Get graph-augmented token-level embeddings \n9 $\\mathbf { H } _ { g } ^ { l + 1 } \\mathrm { T R M } ^ { l } ( \\widehat { \\mathbf { H } } _ { g } ^ { l } )$ ; // Text encoding in Transformer ",
|
| 412 |
+
"bbox": [
|
| 413 |
+
160,
|
| 414 |
+
112,
|
| 415 |
+
730,
|
| 416 |
+
299
|
| 417 |
+
],
|
| 418 |
+
"page_idx": 4
|
| 419 |
+
},
|
| 420 |
+
{
|
| 421 |
+
"type": "text",
|
| 422 |
+
"text": "Return $\\mathbf { h } _ { x } \\mathbf { z } _ { x } ^ { L }$ ; ",
|
| 423 |
+
"bbox": [
|
| 424 |
+
184,
|
| 425 |
+
295,
|
| 426 |
+
318,
|
| 427 |
+
311
|
| 428 |
+
],
|
| 429 |
+
"page_idx": 4
|
| 430 |
+
},
|
| 431 |
+
{
|
| 432 |
+
"type": "text",
|
| 433 |
+
"text": "processed by the layerwise GNN component. The output node-level embeddings $\\hat { \\mathbf { Z } } _ { G } ^ { l }$ are dispatched to their original nodes, which generates the graph-augmented token-level embeddings $\\widehat { \\mathbf { H } } _ { g } ^ { l }$ . The graphaugmented token-level embedding are further processed by the Transformer component. Finally, The node-level embedding (for the center node $x$ ) in the last layer $\\mathbf { z } _ { x } ^ { L }$ is taken as our representation result. ",
|
| 434 |
+
"bbox": [
|
| 435 |
+
173,
|
| 436 |
+
335,
|
| 437 |
+
825,
|
| 438 |
+
396
|
| 439 |
+
],
|
| 440 |
+
"page_idx": 4
|
| 441 |
+
},
|
| 442 |
+
{
|
| 443 |
+
"type": "text",
|
| 444 |
+
"text": "• Encoding Complexity. Given an input of $M$ nodes, each one has $P$ tokens; the time complexity of each layer’s encoding operation is $O ( \\bar { M } ^ { 2 } + M P ^ { 2 } )$ : the graph aggregation takes $O ( M ^ { 2 } )$ , because $M$ node-level embeddings are gathered for multi-head attention; the text encoding takes $\\mathrm { \\Delta } O ( M P ^ { 2 } )$ , as each of the $M$ node calls for the multi-head attention of $P$ tokens. Compared with Transformers, the GNN’s computation cost is much smaller, mainly because of two reasons: 1) $M ^ { 2 } \\ll M P ^ { 2 }$ in general, 2) operations like MLP are not needed in graph aggregation. Therefore, the working efficiency of GraphFormers is close to the cascaded GNN-Transformers as the extra computation cost of layerwise graph aggregation is relatively small. Such a property is also empirically verified in our experiment. ",
|
| 445 |
+
"bbox": [
|
| 446 |
+
173,
|
| 447 |
+
401,
|
| 448 |
+
826,
|
| 449 |
+
513
|
| 450 |
+
],
|
| 451 |
+
"page_idx": 4
|
| 452 |
+
},
|
| 453 |
+
{
|
| 454 |
+
"type": "text",
|
| 455 |
+
"text": "3.2 Model Simplification: Unidirectional Graph Aggregation ",
|
| 456 |
+
"text_level": 1,
|
| 457 |
+
"bbox": [
|
| 458 |
+
174,
|
| 459 |
+
532,
|
| 460 |
+
609,
|
| 461 |
+
549
|
| 462 |
+
],
|
| 463 |
+
"page_idx": 4
|
| 464 |
+
},
|
| 465 |
+
{
|
| 466 |
+
"type": "text",
|
| 467 |
+
"text": "One concern about GraphFormers is that the input nodes are mutually dependent on each other during the encoding process. As a result, to generate the embedding for a node, all the related nodes in its neighbourhood need to be encoded from scratch, regardless of whether they have been processed before. Such a property is unfavorable in practice as a great deal of unnecessary computation cost might be incurred (i.e., a node will be repetitively encoded every time it serves as a neighbour node). We leverage a simple but effective simplification, the unidirectional graph aggregation, to address this problem. Particularly, only the center node $x$ is required to make reference to the neighbourhood; while the rest of nodes $N _ { x }$ remain independently encoded all by their own textual features: ",
|
| 468 |
+
"bbox": [
|
| 469 |
+
173,
|
| 470 |
+
559,
|
| 471 |
+
826,
|
| 472 |
+
671
|
| 473 |
+
],
|
| 474 |
+
"page_idx": 4
|
| 475 |
+
},
|
| 476 |
+
{
|
| 477 |
+
"type": "equation",
|
| 478 |
+
"img_path": "images/eab4f63a052a2ca89da37fa443af75d0e24fd6126aa3ec28fceb51cb1545b746.jpg",
|
| 479 |
+
"text": "$$\n\\mathbf { H } _ { g } ^ { l + 1 } = \\left\\{ \\mathrm { T R M } ^ { l } ( \\widehat { \\mathbf { H } _ { x } ^ { l } } ) , g = x ; \\right.\n$$",
|
| 480 |
+
"text_format": "latex",
|
| 481 |
+
"bbox": [
|
| 482 |
+
375,
|
| 483 |
+
681,
|
| 484 |
+
604,
|
| 485 |
+
724
|
| 486 |
+
],
|
| 487 |
+
"page_idx": 4
|
| 488 |
+
},
|
| 489 |
+
{
|
| 490 |
+
"type": "text",
|
| 491 |
+
"text": "Because the encoding of the neighbour nodes is independent of the center node, the intermediate encoding results $\\{ \\mathbf { z } _ { g } ^ { 1 . . . L } \\} _ { N _ { x } }$ can be cached in storage2 and reused in subsequent computations when they are needed. As a result, the nodes can be prevented from being encoded repetitively, which saves a great deal of unnecessary computation cost. We empirically verify that GraphFormers maintain similar performances when the above simplification is introduced. ",
|
| 492 |
+
"bbox": [
|
| 493 |
+
173,
|
| 494 |
+
733,
|
| 495 |
+
825,
|
| 496 |
+
804
|
| 497 |
+
],
|
| 498 |
+
"page_idx": 4
|
| 499 |
+
},
|
| 500 |
+
{
|
| 501 |
+
"type": "text",
|
| 502 |
+
"text": "3.3 Model Training: Two-Stage Progressive Learning ",
|
| 503 |
+
"text_level": 1,
|
| 504 |
+
"bbox": [
|
| 505 |
+
173,
|
| 506 |
+
823,
|
| 507 |
+
558,
|
| 508 |
+
839
|
| 509 |
+
],
|
| 510 |
+
"page_idx": 4
|
| 511 |
+
},
|
| 512 |
+
{
|
| 513 |
+
"type": "text",
|
| 514 |
+
"text": "• Training Objective. We take advantage of link prediction as our training task. Given a pair of nodes $q$ and $k$ , the model is learned to predict whether they are connected based on their embedding ",
|
| 515 |
+
"bbox": [
|
| 516 |
+
174,
|
| 517 |
+
851,
|
| 518 |
+
823,
|
| 519 |
+
880
|
| 520 |
+
],
|
| 521 |
+
"page_idx": 4
|
| 522 |
+
},
|
| 523 |
+
{
|
| 524 |
+
"type": "table",
|
| 525 |
+
"img_path": "images/c73b9cf82b2b3297561b11d35fe41bdc3b21295297a9a4224ddfc9da4d130a6d.jpg",
|
| 526 |
+
"table_caption": [
|
| 527 |
+
"Table 1: Specifications of the experimental datasets: the number of items, the number of neighbour nodes on average, and the number of training, validation, testing cases. "
|
| 528 |
+
],
|
| 529 |
+
"table_footnote": [],
|
| 530 |
+
"table_body": "<table><tr><td></td><td>Product</td><td>DBLP</td><td>Wiki</td></tr><tr><td>#Item</td><td>5,643,688</td><td>4,894,081</td><td>4,818.679</td></tr><tr><td>#N</td><td>4.71</td><td>9.31</td><td>8.86</td></tr><tr><td>#Train</td><td>22,146,934</td><td>3,009,506</td><td>7,145,834</td></tr><tr><td>#Valid</td><td>30,000</td><td>60,000</td><td>66,167</td></tr><tr><td>#Test</td><td>306,742</td><td>100,000</td><td>100,000</td></tr></table>",
|
| 531 |
+
"bbox": [
|
| 532 |
+
336,
|
| 533 |
+
125,
|
| 534 |
+
658,
|
| 535 |
+
202
|
| 536 |
+
],
|
| 537 |
+
"page_idx": 5
|
| 538 |
+
},
|
| 539 |
+
{
|
| 540 |
+
"type": "text",
|
| 541 |
+
"text": "similarity. Particularly, the following classification loss is minimized for a positive pair of $q$ and $k$ ",
|
| 542 |
+
"bbox": [
|
| 543 |
+
176,
|
| 544 |
+
214,
|
| 545 |
+
815,
|
| 546 |
+
229
|
| 547 |
+
],
|
| 548 |
+
"page_idx": 5
|
| 549 |
+
},
|
| 550 |
+
{
|
| 551 |
+
"type": "equation",
|
| 552 |
+
"img_path": "images/e8bd68d70566214ab8a3c6da385683ac439ffa7a92236be94eb21e5c96fe083b.jpg",
|
| 553 |
+
"text": "$$\n\\mathcal { L } = - \\log \\frac { \\exp ( \\langle \\mathbf { h } _ { q } , \\mathbf { h } _ { k } \\rangle ) } { \\exp ( \\langle \\mathbf { h } _ { q } , \\mathbf { h } _ { k } \\rangle ) + \\sum _ { r \\in R } \\exp ( \\langle \\mathbf { h } _ { q } , \\mathbf { h } _ { r } \\rangle ) } .\n$$",
|
| 554 |
+
"text_format": "latex",
|
| 555 |
+
"bbox": [
|
| 556 |
+
330,
|
| 557 |
+
231,
|
| 558 |
+
666,
|
| 559 |
+
267
|
| 560 |
+
],
|
| 561 |
+
"page_idx": 5
|
| 562 |
+
},
|
| 563 |
+
{
|
| 564 |
+
"type": "text",
|
| 565 |
+
"text": "In the above equation, $\\mathbf { h } _ { q }$ and $\\mathbf { h } _ { k }$ are the node embeddings; $\\langle \\cdot \\rangle$ denotes the computation of inner product; $R$ stands for the negative samples. In our implementation, we leverage “in-batch negative samples” (Karpukhin et al., 2020; Luan et al., 2020) for the reduction of encoding cost: a positive sample in one training instance will be used as a negative sample in the rest of the training instances within the same mini-batch. ",
|
| 566 |
+
"bbox": [
|
| 567 |
+
173,
|
| 568 |
+
270,
|
| 569 |
+
825,
|
| 570 |
+
339
|
| 571 |
+
],
|
| 572 |
+
"page_idx": 5
|
| 573 |
+
},
|
| 574 |
+
{
|
| 575 |
+
"type": "text",
|
| 576 |
+
"text": "• Two-stage Training. In GraphFormers, the information from the center node and neighbour nodes are not treated equally, which may undermine the model’s training effect. Particularly, the center node’s information can be directly utilized, while the neighbourhood information needs to be introduced via three steps: 1) encoded as node-level embeddings, 2) making graph aggregation with the center node, and 3) introduced to center node’s graph augmented token-level embeddings. The message passing pathway can shortcut when the center nodes are “sufficiently informative”, i.e., two nodes are sufficiently similar with each other in terms of their own textual features, such that their connection can be predicted without considering the neighbours. Given the existence of such cases, GraphFormers may end up with well-trained Transformers but insufficiently trained GNNs. ",
|
| 577 |
+
"bbox": [
|
| 578 |
+
173,
|
| 579 |
+
344,
|
| 580 |
+
825,
|
| 581 |
+
470
|
| 582 |
+
],
|
| 583 |
+
"page_idx": 5
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"type": "text",
|
| 587 |
+
"text": "To alleviate the above problem, we introduce a warm-up training task, where the link prediction is made based on the polluted input nodes. Particularly, for each input node $g$ , a subset of its tokens $g _ { m }$ will be randomly masked4. As a result, the classification loss becomes: ",
|
| 588 |
+
"bbox": [
|
| 589 |
+
174,
|
| 590 |
+
477,
|
| 591 |
+
821,
|
| 592 |
+
518
|
| 593 |
+
],
|
| 594 |
+
"page_idx": 5
|
| 595 |
+
},
|
| 596 |
+
{
|
| 597 |
+
"type": "equation",
|
| 598 |
+
"img_path": "images/762c25ff6c3e2572e827b1ca9794b136d56ef00b5a836e29cf832b028987367a.jpg",
|
| 599 |
+
"text": "$$\n\\mathcal { L } ^ { \\prime } = - \\log \\frac { \\exp ( \\langle \\mathbf { h } _ { \\widetilde { q } } , \\mathbf { h } _ { \\widetilde { k } } \\rangle ) } { \\exp ( \\langle \\mathbf { h } _ { \\widetilde { q } } , \\mathbf { h } _ { \\widetilde { k } } \\rangle ) + \\sum _ { r \\in R } \\exp ( \\langle \\mathbf { h } _ { \\widetilde { q } } , \\mathbf { h } _ { \\widetilde { r } } \\rangle ) } ,\n$$",
|
| 600 |
+
"text_format": "latex",
|
| 601 |
+
"bbox": [
|
| 602 |
+
328,
|
| 603 |
+
521,
|
| 604 |
+
668,
|
| 605 |
+
558
|
| 606 |
+
],
|
| 607 |
+
"page_idx": 5
|
| 608 |
+
},
|
| 609 |
+
{
|
| 610 |
+
"type": "text",
|
| 611 |
+
"text": "where $\\mathbf { h } _ { \\widetilde { q } } , \\mathbf { h } _ { \\widetilde { k } } .$ , $\\mathbf { h } _ { \\tilde { r } }$ are the embeddings generated from the polluted nodes. The masked tokens reduce the informativeness of each individual node; therefore, the model is forced to leverage the whole input nodes to make the right prediction. ",
|
| 612 |
+
"bbox": [
|
| 613 |
+
173,
|
| 614 |
+
559,
|
| 615 |
+
825,
|
| 616 |
+
602
|
| 617 |
+
],
|
| 618 |
+
"page_idx": 5
|
| 619 |
+
},
|
| 620 |
+
{
|
| 621 |
+
"type": "text",
|
| 622 |
+
"text": "Finally, the model training is organized as a two-stage progressive learning process. In the first stage, the model is trained to minimize $\\mathcal { L } ^ { \\prime }$ based on the polluted nodes until its convergence, which reinforce the model’s capability of integrating information on graph. In the second stage, the model is continually trained to minimize $\\mathcal { L }$ based on the original data until the convergence, which makes the model fit into the target distribution. ",
|
| 623 |
+
"bbox": [
|
| 624 |
+
174,
|
| 625 |
+
607,
|
| 626 |
+
825,
|
| 627 |
+
678
|
| 628 |
+
],
|
| 629 |
+
"page_idx": 5
|
| 630 |
+
},
|
| 631 |
+
{
|
| 632 |
+
"type": "text",
|
| 633 |
+
"text": "4 Experimental Studies ",
|
| 634 |
+
"text_level": 1,
|
| 635 |
+
"bbox": [
|
| 636 |
+
174,
|
| 637 |
+
695,
|
| 638 |
+
385,
|
| 639 |
+
713
|
| 640 |
+
],
|
| 641 |
+
"page_idx": 5
|
| 642 |
+
},
|
| 643 |
+
{
|
| 644 |
+
"type": "text",
|
| 645 |
+
"text": "4.1 Data and Settings ",
|
| 646 |
+
"text_level": 1,
|
| 647 |
+
"bbox": [
|
| 648 |
+
174,
|
| 649 |
+
726,
|
| 650 |
+
336,
|
| 651 |
+
741
|
| 652 |
+
],
|
| 653 |
+
"page_idx": 5
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"type": "text",
|
| 657 |
+
"text": "We make use of the following three real-world textual graph datasets for our experimental studies. ",
|
| 658 |
+
"bbox": [
|
| 659 |
+
173,
|
| 660 |
+
752,
|
| 661 |
+
812,
|
| 662 |
+
767
|
| 663 |
+
],
|
| 664 |
+
"page_idx": 5
|
| 665 |
+
},
|
| 666 |
+
{
|
| 667 |
+
"type": "text",
|
| 668 |
+
"text": "• $\\mathbf { \\nabla } \\mathbf { D B L P } ^ { 5 }$ , which contains the paper citation graph from DBLP up to 2020-04-09. Two papers are linked if one is cited by the other one. The paper’s title is used as the textual feature. ",
|
| 669 |
+
"bbox": [
|
| 670 |
+
174,
|
| 671 |
+
772,
|
| 672 |
+
821,
|
| 673 |
+
801
|
| 674 |
+
],
|
| 675 |
+
"page_idx": 5
|
| 676 |
+
},
|
| 677 |
+
{
|
| 678 |
+
"type": "text",
|
| 679 |
+
"text": "• Wikidata5M6 (Wiki) (Wang et al., 2019b), which contains the entity graph from Wikipedia. The first sentence in each entity’s introduction is taken as its textual feature. ",
|
| 680 |
+
"bbox": [
|
| 681 |
+
173,
|
| 682 |
+
806,
|
| 683 |
+
821,
|
| 684 |
+
835
|
| 685 |
+
],
|
| 686 |
+
"page_idx": 5
|
| 687 |
+
},
|
| 688 |
+
{
|
| 689 |
+
"type": "table",
|
| 690 |
+
"img_path": "images/f52a662686793232cb4ee4b0aa143c90541982b6bc46e3393f69a6278c18457e.jpg",
|
| 691 |
+
"table_caption": [
|
| 692 |
+
"Table 2: Overall evaluation (GraphFormers marked in bold, the best baseline underlined). GraphFormers outperforms all baselines, especially the ones based on cascaded Transformers-GNN. "
|
| 693 |
+
],
|
| 694 |
+
"table_footnote": [],
|
| 695 |
+
"table_body": "<table><tr><td></td><td colspan=\"3\">Product</td><td colspan=\"3\">DBLP</td><td colspan=\"3\">Wiki</td></tr><tr><td>Methods</td><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td></tr><tr><td>PLM</td><td>0.6563</td><td>0.7911</td><td>0.7344</td><td>0.5673</td><td>0.7484</td><td>0.6777</td><td>0.3466</td><td>0.5799</td><td>0.4712</td></tr><tr><td>TNVE</td><td>0.4618</td><td>0.6204</td><td>0.5364</td><td>0.2978</td><td>0.5295</td><td>0.4163</td><td>0.1786</td><td>0.4274</td><td>0.2933</td></tr><tr><td>IFTN</td><td>0.5233</td><td>0.6740</td><td>0.5982</td><td>0.3691</td><td>0.5798</td><td>0.4773</td><td>0.1838</td><td>0.4276</td><td>0.2945</td></tr><tr><td>PLM+GAT</td><td>0.7540</td><td>0.8637</td><td>0.8232</td><td>0.6633</td><td>0.8204</td><td>0.7667</td><td>0.3006</td><td>0.5430</td><td>0.4270</td></tr><tr><td>PLM+Max</td><td>0.7570</td><td>0.8678</td><td>0.8280</td><td>0.6934</td><td>0.8386</td><td>0.7900</td><td>0.3712</td><td>0.6071</td><td>0.5022</td></tr><tr><td>PLM+Mean</td><td>0.7550</td><td>0.8671</td><td>0.8271</td><td>0.6896</td><td>0.8359</td><td>0.7866</td><td>0.3664</td><td>0.6037</td><td>0.4980</td></tr><tr><td>PLM+Att</td><td>0.7513</td><td>0.8652</td><td>0.8246</td><td>0.6910</td><td>0.8366</td><td>0.7875</td><td>0.3709</td><td>0.6067</td><td>0.5018</td></tr><tr><td>GraphFormers</td><td>0.7786</td><td>0.8793</td><td>0.8430</td><td>0.7267</td><td>0.8565</td><td>0.8133</td><td>0.3952</td><td>0.6230</td><td>0.5220</td></tr></table>",
|
| 696 |
+
"bbox": [
|
| 697 |
+
173,
|
| 698 |
+
123,
|
| 699 |
+
826,
|
| 700 |
+
256
|
| 701 |
+
],
|
| 702 |
+
"page_idx": 6
|
| 703 |
+
},
|
| 704 |
+
{
|
| 705 |
+
"type": "text",
|
| 706 |
+
"text": "• Product Graph (Product), an even larger dataset of online products collected by a world-wide search engine. In this dataset, the users’ web browsing behaviors are tracked for the targeted product webpages (e.g., Amazon webpages of Nike shoes). The user’s continuously browsed webpages within a short period of time (e.g., 30 minutes) is called a “session”. The products within a common session are connected in the graph (which is a common way of graph construction in e-commerce scenarios (Ying et al., 2018; Wang et al., 2018)). Each product has its unique textual description, which specifies information like the product name, brand, and saler, etc. ",
|
| 707 |
+
"bbox": [
|
| 708 |
+
173,
|
| 709 |
+
270,
|
| 710 |
+
825,
|
| 711 |
+
367
|
| 712 |
+
],
|
| 713 |
+
"page_idx": 6
|
| 714 |
+
},
|
| 715 |
+
{
|
| 716 |
+
"type": "text",
|
| 717 |
+
"text": "The textual features of all the datasets are in English. We make use of uncased WordPiece (Wu et al., 2016) to tokenize the input text. In our experiment, each text is associated with 5 uniformly sampled neighbours (without replacement); for texts with neighbourhood smaller than 5, all the neighbours will be utilized. We summarized the specifications of all the datasets with Table 1. The experiment results are evaluated in terms of link prediction accuracy, i.e., to predict whether a query node and key node are connected given the textual features of themselves and their neighbours. In each testing instance, one query is provided with 300 keys: 1 positive plus 299 randomly sampled negative cases. We leverage three common metrics to measure the prediction accuracy: Precision $@ 1$ , NDCG, and MRR. Without specifications, we will take the unidirectional-simplified GraphFormers trained with the two-stage progressive learning as our default model. More details about the implementations and the training/testing configurations are summarized in an Appendix file. It is submitted together with our source code within the supplementary materials. ",
|
| 718 |
+
"bbox": [
|
| 719 |
+
173,
|
| 720 |
+
372,
|
| 721 |
+
826,
|
| 722 |
+
539
|
| 723 |
+
],
|
| 724 |
+
"page_idx": 6
|
| 725 |
+
},
|
| 726 |
+
{
|
| 727 |
+
"type": "text",
|
| 728 |
+
"text": "4.2 Baselines ",
|
| 729 |
+
"text_level": 1,
|
| 730 |
+
"bbox": [
|
| 731 |
+
174,
|
| 732 |
+
564,
|
| 733 |
+
276,
|
| 734 |
+
579
|
| 735 |
+
],
|
| 736 |
+
"page_idx": 6
|
| 737 |
+
},
|
| 738 |
+
{
|
| 739 |
+
"type": "text",
|
| 740 |
+
"text": "We focus on the comparison between GNN-nested Transformers and Cascaded Transformers-GNN. To make sure the difference between both architectures can be truthfully reflected from the evaluation results, GraphFormers and the Cascaded Transformers-GNN baselines are equipped with text encoders and graph aggregators of the same capacities. Particularly, we use the BERT-like PLM as our text encoder, where UniLM-base7 (Bao et al., 2020) is chosen as the network backbone for all related methods; the final layer’s [CLS] token embedding is used for the text embedding. ",
|
| 741 |
+
"bbox": [
|
| 742 |
+
174,
|
| 743 |
+
594,
|
| 744 |
+
825,
|
| 745 |
+
678
|
| 746 |
+
],
|
| 747 |
+
"page_idx": 6
|
| 748 |
+
},
|
| 749 |
+
{
|
| 750 |
+
"type": "text",
|
| 751 |
+
"text": "We enumerate the following representative graph aggregators as used in GAT (Velickovi ˇ c et al., 2018), ´ GIN (Xu et al., 2018), GraphSage (Hamilton et al., 2017a). The GAT aggregator, where the node embedding is generated as the weighted sum of all the text embeddings. Each text embedding’s relative importance is calculated as the attention score with the center node. The Pooling-andConcat aggregators, where the center node’s text embedding is concatenated with the neighbours’ pooling result and linearly transformed for the final representation. Depending on the form of pooling function, we have the following options: Max and Mean, where neighbours are aggregated by max-pooling and mean-pooling, respectively; Att, where the neighbours are summed up based on the attention weights with the center node. By comparison, the neighbourhood information may get more emphasized with GAT; while the center node itself tends to be highlighted with Pooling-and-Concat. ",
|
| 752 |
+
"bbox": [
|
| 753 |
+
173,
|
| 754 |
+
684,
|
| 755 |
+
825,
|
| 756 |
+
821
|
| 757 |
+
],
|
| 758 |
+
"page_idx": 6
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"type": "text",
|
| 762 |
+
"text": "We consider two more baselines which make use of simplified text encoders (such as CNN) and network embeddings: TNVE (Wang et al., 2019a) and IFTN (Xu et al., 2019). We also include the PLM only baseline, which merely leverages the textual feature of the center node. ",
|
| 763 |
+
"bbox": [
|
| 764 |
+
176,
|
| 765 |
+
828,
|
| 766 |
+
821,
|
| 767 |
+
871
|
| 768 |
+
],
|
| 769 |
+
"page_idx": 6
|
| 770 |
+
},
|
| 771 |
+
{
|
| 772 |
+
"type": "table",
|
| 773 |
+
"img_path": "images/3cab5fc3e17d2e70b4f3ad08877ef844e07e0982e0e7bcf50a70e2f57fccc951.jpg",
|
| 774 |
+
"table_caption": [
|
| 775 |
+
"Table 3: Impact of neighbour size (#N). "
|
| 776 |
+
],
|
| 777 |
+
"table_footnote": [],
|
| 778 |
+
"table_body": "<table><tr><td rowspan=\"2\">#N</td><td colspan=\"3\">GraphFormers</td><td colspan=\"3\">PLM+Max</td></tr><tr><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td></tr><tr><td>1</td><td>0.6485</td><td>0.8087</td><td>0.7522</td><td>0.6249</td><td>0.7946</td><td>0.7342</td></tr><tr><td>2</td><td>0.6841</td><td>0.8308</td><td>0.7804</td><td>0.6538</td><td>0.8137</td><td>0.7583</td></tr><tr><td>3</td><td>0.6980</td><td>0.8396</td><td>0.7916</td><td>0.6728</td><td>0.8256</td><td>0.7734</td></tr><tr><td>4</td><td>0.7126</td><td>0.8485</td><td>0.8029</td><td>0.6823</td><td>0.8319</td><td>0.7814</td></tr><tr><td>5</td><td>0.7267</td><td>0.8565</td><td>0.8133</td><td>0.6934</td><td>0.8386</td><td>0.7900</td></tr></table>",
|
| 779 |
+
"bbox": [
|
| 780 |
+
326,
|
| 781 |
+
111,
|
| 782 |
+
669,
|
| 783 |
+
207
|
| 784 |
+
],
|
| 785 |
+
"page_idx": 7
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"type": "text",
|
| 789 |
+
"text": "4.3 Overall Evaluation ",
|
| 790 |
+
"text_level": 1,
|
| 791 |
+
"bbox": [
|
| 792 |
+
174,
|
| 793 |
+
218,
|
| 794 |
+
344,
|
| 795 |
+
233
|
| 796 |
+
],
|
| 797 |
+
"page_idx": 7
|
| 798 |
+
},
|
| 799 |
+
{
|
| 800 |
+
"type": "text",
|
| 801 |
+
"text": "The overall evaluation results are reported in Table 2. It’s observed that GraphFormers consistently outperform all the baselines, especially the ones based on the cascaded Transformers-GNN, with notable advantages. Particularly, it achieves $2 . 9 \\%$ , $4 . 8 \\%$ , $6 . 5 \\%$ relative improvements over the most competitive baselines (underlined) on each of the experimental datasets. Such an observation indicates that the relationship between the nodes can be captured more accurately based on the node embeddings generated by GraphFormers, which verifies the effectiveness of our proposed method. ",
|
| 802 |
+
"bbox": [
|
| 803 |
+
174,
|
| 804 |
+
256,
|
| 805 |
+
825,
|
| 806 |
+
339
|
| 807 |
+
],
|
| 808 |
+
"page_idx": 7
|
| 809 |
+
},
|
| 810 |
+
{
|
| 811 |
+
"type": "text",
|
| 812 |
+
"text": "We also observe the following underlying factors that may influence the representation quality. ",
|
| 813 |
+
"bbox": [
|
| 814 |
+
174,
|
| 815 |
+
345,
|
| 816 |
+
790,
|
| 817 |
+
361
|
| 818 |
+
],
|
| 819 |
+
"page_idx": 7
|
| 820 |
+
},
|
| 821 |
+
{
|
| 822 |
+
"type": "text",
|
| 823 |
+
"text": "Firstly, the effective utilization of neighbourhood information is critical. With the joint consideration of the center node and neighbour nodes, the $\\mathrm { P L M + G N N s }$ methods, including GraphFormers and the cascaded Transformers-GNN baselines, significantly outperform the PLM only baseline in most of the time. We further analyze the impact of neighbourhood size as Table 3, with a fraction of neighbour nodes randomly sampled for each center node (using DBLP for illustration). It can be observed that both GraphFormers and $\\mathrm { P L M + M a x }$ (the most competitive baseline) achieve higher prediction accuracy than the PLM only method $( \\mathrm { P } \\ @ 1 . 0 . 5 6 7 3$ , NDCG:0.7484, MRR:0.6777, as reported in Table 2), even with fewer neighbour nodes included. With the increasing number of neighbour nodes, the advantages become gradually enlarged. However, the marginal gain is vanishing, as the relative improvement becomes smaller when more neighbours are included. In all the testing cases, GraphFormers maintain consistent advantages over PLM+Max, which reaffirms the effectiveness of our proposed methods. ",
|
| 824 |
+
"bbox": [
|
| 825 |
+
174,
|
| 826 |
+
366,
|
| 827 |
+
825,
|
| 828 |
+
531
|
| 829 |
+
],
|
| 830 |
+
"page_idx": 7
|
| 831 |
+
},
|
| 832 |
+
{
|
| 833 |
+
"type": "text",
|
| 834 |
+
"text": "Secondly, the capacity of the text encoder is crucial for textual graph representation. All the pretrained language model based methods (GraphFormers, Cascaded Transformers-GNN baselines, PLM-only baseline) significantly outperform the baselines with simplified text encoders (TNVE, IFTN). ",
|
| 835 |
+
"bbox": [
|
| 836 |
+
176,
|
| 837 |
+
539,
|
| 838 |
+
823,
|
| 839 |
+
580
|
| 840 |
+
],
|
| 841 |
+
"page_idx": 7
|
| 842 |
+
},
|
| 843 |
+
{
|
| 844 |
+
"type": "text",
|
| 845 |
+
"text": "Thirdly, the representation quality is also sensitive to the form of graph aggregator. In Product, the cascaded Transformers-GNN baselines’ performances are quite close to each other. In DBLP, $\\mathrm { P L M + }$ (Max, Mean, Att) outperforms PLM+GAT. In Wiki, not only $\\mathrm { P L M + }$ (Max, Mean, Att) but also PLM-only baseline outperform PLM $^ +$ GAT. Such phenomenons could be attributed to the type of graph: whether it is homogeneous or heterogeneous. Particularly, both Product and DBLP can be regarded as homogeneous graphs as the nodes are connected based on the same relationships; i.e., co-view relationship in Product, and citation relationship in DBLP. In both homogeneous graphs, the connected nodes may have quite similar semantics (the co-viewed products usually serve similar user intents, and the citation relationships usually indicate similar research topics); thus, the incorporated neighbour nodes will probably provide complementary information for the link prediction between the center nodes. However, Wiki is a heterogeneous graph, where the connections between entities may have highly different semantics. As a result, the incorporation of neighbour nodes may not contribute to the link prediction task, especially when the incorporated neighbours and the prediction target are connected to the center nodes with totally different relationships. Considering that GAT tends to focus more on the neighbourhood, its performance can be vulnerable in such unfavorable situations. These findings suggest that the neighbourhood information should be properly handled in case that the information of the center node is wiped out. ",
|
| 846 |
+
"bbox": [
|
| 847 |
+
174,
|
| 848 |
+
587,
|
| 849 |
+
825,
|
| 850 |
+
821
|
| 851 |
+
],
|
| 852 |
+
"page_idx": 7
|
| 853 |
+
},
|
| 854 |
+
{
|
| 855 |
+
"type": "text",
|
| 856 |
+
"text": "Finally, we may conclude different methods’ utility in textual graph representation: simplified text encoders $\\prec P L M s \\prec C$ ascaded Transformers-GNN $\\prec$ GNNs-nested Transformers. Such findings are consistent with our expectation that the precise modeling of individual textual feature and the effective integration of neighbourhood information will jointly contribute to high-quality textual graph representation. GraphFormers enjoy the high expressiveness of PLMs and leverage layerwise nested-GNNs to facilitate graph aggregation, which contributes to both of the above perspectives. ",
|
| 857 |
+
"bbox": [
|
| 858 |
+
174,
|
| 859 |
+
828,
|
| 860 |
+
823,
|
| 861 |
+
911
|
| 862 |
+
],
|
| 863 |
+
"page_idx": 7
|
| 864 |
+
},
|
| 865 |
+
{
|
| 866 |
+
"type": "table",
|
| 867 |
+
"img_path": "images/43c1354e1bfe1d97b2a6ef35a0d8888712760e3823b7cca5716f7baa6f6aa5ad.jpg",
|
| 868 |
+
"table_caption": [
|
| 869 |
+
"Table 4: Ablation Studies (The top ablated methods are marked in bold; $\\uparrow ^ { \\mathrm { , , } } / \\downarrow ^ { \\mathrm { , } } \\downarrow ^ { \\mathrm { , } }$ : the performance is increased/decreased compared with the default setting). “-Progressive”: two-stage progressive learning disabled; “-Simplified”: unidirectional simplification disabled; “-Shared GNNs”: GNNs parameters are not shared across the layers; “-Position”: GNNs learnable position bias disabled. "
|
| 870 |
+
],
|
| 871 |
+
"table_footnote": [],
|
| 872 |
+
"table_body": "<table><tr><td></td><td colspan=\"3\">Product</td><td colspan=\"3\">DBLP</td><td colspan=\"3\">Wiki</td></tr><tr><td>Methods</td><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td><td>P@1</td><td>NDCG</td><td>MRR</td></tr><tr><td>GraphFormers</td><td>0.7786</td><td>0.8793</td><td>0.8430</td><td>0.7267</td><td>0.8565</td><td>0.8133</td><td>0.3952</td><td>0.6230</td><td>0.5220</td></tr><tr><td>PLM+Max</td><td>0.7570</td><td>0.8678</td><td>0.8280</td><td>0.6934</td><td>0.8386</td><td>0.7900</td><td>0.3712</td><td>0.6071</td><td>0.5022</td></tr><tr><td>- Progressive</td><td>0.7688</td><td>0.8751</td><td>0.8373</td><td>0.7096</td><td>0.8468</td><td>0.8007</td><td>0.3834</td><td>0.6155</td><td>0.5127</td></tr><tr><td>- Simplified</td><td>0.7795个</td><td>0.8798 个</td><td>0.8436 个</td><td>0.7225</td><td>0.8542</td><td>0.8102</td><td>0.3923</td><td>0.6209</td><td>0.5195</td></tr><tr><td>- Shared GNNs</td><td>0.7788</td><td>0.8795</td><td>0.8433</td><td>0.7256</td><td>0.8558</td><td>0.8123</td><td>0.3945 ↓</td><td>0.6221↓</td><td>0.5211 ↓</td></tr><tr><td>- Position</td><td>0.7788</td><td>0.8795</td><td>0.8434</td><td>0.7276个</td><td>0.8570个</td><td>0.8139个</td><td>0.3942</td><td>0.6222</td><td>0.5211</td></tr></table>",
|
| 873 |
+
"bbox": [
|
| 874 |
+
176,
|
| 875 |
+
151,
|
| 876 |
+
821,
|
| 877 |
+
267
|
| 878 |
+
],
|
| 879 |
+
"page_idx": 8
|
| 880 |
+
},
|
| 881 |
+
{
|
| 882 |
+
"type": "table",
|
| 883 |
+
"img_path": "images/004dc65c93f2cbdf2688479d8e34ce91187f43ac6b135de15b00f76fcafa696a.jpg",
|
| 884 |
+
"table_caption": [
|
| 885 |
+
"Table 5: Time and memory costs per mini-batch for $\\mathrm { P L M + M a x }$ and GraphFormers, with neighbour size increased from 3 to 200. GraphFormers achieve similar efficiency and scalability as $\\mathrm { P L M + M a x }$ "
|
| 886 |
+
],
|
| 887 |
+
"table_footnote": [],
|
| 888 |
+
"table_body": "<table><tr><td>#N</td><td>3</td><td>5</td><td>10</td><td>20</td><td>50</td><td>100</td><td>200</td></tr><tr><td>Time: PLM+Max</td><td>60.29 ms</td><td>93.41 ms</td><td>161.40 ms</td><td>295.92 ms</td><td>684.16 ms</td><td>1357.93 ms</td><td>2706.35 ms</td></tr><tr><td>Time: GraphFormers</td><td>63.95 ms</td><td>97.19 ms</td><td>170.16 ms</td><td>306.12 ms</td><td>714.32 ms</td><td>1411.09 ms</td><td>2801.67 ms</td></tr><tr><td>Mem: PLM+Max</td><td>1.33 GiB</td><td>1.39 GiB</td><td>1.55 GiB</td><td>1.82 GiB</td><td>2.67 GiB</td><td>4.09 GiB</td><td>6.92 GiB</td></tr><tr><td>Mem: GraphFormers</td><td>1.33 GiB</td><td>1.39 GiB</td><td>1.55 GiB</td><td>1.83 GiB</td><td>2.70 GiB</td><td>4.28 GiB</td><td>7.33 GiB</td></tr></table>",
|
| 889 |
+
"bbox": [
|
| 890 |
+
173,
|
| 891 |
+
319,
|
| 892 |
+
825,
|
| 893 |
+
390
|
| 894 |
+
],
|
| 895 |
+
"page_idx": 8
|
| 896 |
+
},
|
| 897 |
+
{
|
| 898 |
+
"type": "text",
|
| 899 |
+
"text": "4.4 Ablation Studies ",
|
| 900 |
+
"text_level": 1,
|
| 901 |
+
"bbox": [
|
| 902 |
+
174,
|
| 903 |
+
404,
|
| 904 |
+
328,
|
| 905 |
+
419
|
| 906 |
+
],
|
| 907 |
+
"page_idx": 8
|
| 908 |
+
},
|
| 909 |
+
{
|
| 910 |
+
"type": "text",
|
| 911 |
+
"text": "The ablation studies (as Table 4) are performed to clarify the following issues: 1) the impact of two-stage progressive learning, and 2) the impact of unidirectional-simplified GraphFormers. ",
|
| 912 |
+
"bbox": [
|
| 913 |
+
173,
|
| 914 |
+
430,
|
| 915 |
+
823,
|
| 916 |
+
458
|
| 917 |
+
],
|
| 918 |
+
"page_idx": 8
|
| 919 |
+
},
|
| 920 |
+
{
|
| 921 |
+
"type": "text",
|
| 922 |
+
"text": "Firstly, the two-stage progressive learning substantially improves GraphFormers’ representation quality. Without such a training strategy (\"-Progressive\": training directly on the original data), the model’s performance is decreased by $0 . 9 8 \\%$ , $1 . 7 1 \\%$ , and $1 . 1 8 \\%$ in each of the datasets, respectively. ",
|
| 923 |
+
"bbox": [
|
| 924 |
+
176,
|
| 925 |
+
465,
|
| 926 |
+
823,
|
| 927 |
+
507
|
| 928 |
+
],
|
| 929 |
+
"page_idx": 8
|
| 930 |
+
},
|
| 931 |
+
{
|
| 932 |
+
"type": "text",
|
| 933 |
+
"text": "Secondly, the performances between simplified and non-simplified (“-Simplified”) GraphFormers are comparable. In fact, the necessity of graph aggregation is not equivalent for the center node and the neighbour nodes: since the center node is the one for representation, it is much more important to ensure that the center node may extract complementary information from its neighbours. The unidirectional-simplified GraphFormers maintain such a property; thus, there is little impact on the final performances. Such a finding affirms that we may safely leverage the simplified model to save the cost of repetitively encoding the existing neighbours. ",
|
| 934 |
+
"bbox": [
|
| 935 |
+
174,
|
| 936 |
+
512,
|
| 937 |
+
825,
|
| 938 |
+
611
|
| 939 |
+
],
|
| 940 |
+
"page_idx": 8
|
| 941 |
+
},
|
| 942 |
+
{
|
| 943 |
+
"type": "text",
|
| 944 |
+
"text": "We make two additional ablation studies. “-Shared GNNs”: the GNNs parameters sharing is disabled, where each layer maintains its own graph aggregator (by default, the layerwise GNN components in GraphFormers share the same set of parameters). “-Position”: the learnable position bias $\\mathbf { b }$ in Eq. 1) is disabled in GNNs. We find that model’s performance is little affected from the above changes. ",
|
| 945 |
+
"bbox": [
|
| 946 |
+
174,
|
| 947 |
+
616,
|
| 948 |
+
825,
|
| 949 |
+
672
|
| 950 |
+
],
|
| 951 |
+
"page_idx": 8
|
| 952 |
+
},
|
| 953 |
+
{
|
| 954 |
+
"type": "text",
|
| 955 |
+
"text": "4.5 Efficiency Analysis ",
|
| 956 |
+
"text_level": 1,
|
| 957 |
+
"bbox": [
|
| 958 |
+
176,
|
| 959 |
+
690,
|
| 960 |
+
344,
|
| 961 |
+
705
|
| 962 |
+
],
|
| 963 |
+
"page_idx": 8
|
| 964 |
+
},
|
| 965 |
+
{
|
| 966 |
+
"type": "text",
|
| 967 |
+
"text": "We compare the time efficiency between GNN-nested Transformers (GraphFormers) and Cascaded Transformers+GNN (using PLM+Max for comparison). The evaluation is made with a Nvidia P100 GPU. Each mini-batch contains 32 encoding instances; each instance contains one center and #N neighbour nodes; the token length of each node is 16. We report the average time and memory (GPU RAM) costs per mini-batch as Table 5. ",
|
| 968 |
+
"bbox": [
|
| 969 |
+
174,
|
| 970 |
+
717,
|
| 971 |
+
825,
|
| 972 |
+
786
|
| 973 |
+
],
|
| 974 |
+
"page_idx": 8
|
| 975 |
+
},
|
| 976 |
+
{
|
| 977 |
+
"type": "text",
|
| 978 |
+
"text": "Firstly, the time and memory costs of both methods grow linearly with the increment of neighbours. (There are overheads of time and memory costs. The time cost overhead may come from CPU processing; while the memory cost overhead is mainly due to the model parameters (Rajbhandari et al., 2020)). We may approximately remove the overheads by deducting the time and memory costs where $\\# \\mathrm { N } { = } 3$ ). Such a finding is consistent with our theoretical analysis in Section 3.1. ",
|
| 979 |
+
"bbox": [
|
| 980 |
+
174,
|
| 981 |
+
794,
|
| 982 |
+
825,
|
| 983 |
+
863
|
| 984 |
+
],
|
| 985 |
+
"page_idx": 8
|
| 986 |
+
},
|
| 987 |
+
{
|
| 988 |
+
"type": "text",
|
| 989 |
+
"text": "Secondly, the overall time and memory costs of GraphFormers are quite close to $\\mathrm { P L M + M a x }$ . When the number of neighbour nodes is small, the differences between both methods are almost ignorable. The differences become slightly larger when more neighbour nodes are included, because the layerwise graph aggregations in GraphFormers get increasingly time consuming. However, the differences are still relatively small: merely around $3 . 5 \\%$ of the overall running costs when #N is increased to 200 $\\because \\angle A N = 2 0 0 ^ { \\circ }$ is already more than enough for most of the real world scenarios). ",
|
| 990 |
+
"bbox": [
|
| 991 |
+
176,
|
| 992 |
+
869,
|
| 993 |
+
823,
|
| 994 |
+
911
|
| 995 |
+
],
|
| 996 |
+
"page_idx": 8
|
| 997 |
+
},
|
| 998 |
+
{
|
| 999 |
+
"type": "image",
|
| 1000 |
+
"img_path": "images/8d72d1d3cf6982f405391a7c8bc1dc88ff1373034b704b69b86bbd68441b0848.jpg",
|
| 1001 |
+
"image_caption": [
|
| 1002 |
+
"Figure 3: Online $\\mathrm { A } / \\mathrm { B }$ Test: the relative improvements of RPM, CY and CPC against the last version of production system in Bing Search (green: positive; blue: negative). In most of the time, all three performance indicators are significantly improved thanks to the utilization of GraphFormers. "
|
| 1003 |
+
],
|
| 1004 |
+
"image_footnote": [],
|
| 1005 |
+
"bbox": [
|
| 1006 |
+
174,
|
| 1007 |
+
88,
|
| 1008 |
+
823,
|
| 1009 |
+
194
|
| 1010 |
+
],
|
| 1011 |
+
"page_idx": 9
|
| 1012 |
+
},
|
| 1013 |
+
{
|
| 1014 |
+
"type": "text",
|
| 1015 |
+
"text": "",
|
| 1016 |
+
"bbox": [
|
| 1017 |
+
174,
|
| 1018 |
+
261,
|
| 1019 |
+
825,
|
| 1020 |
+
303
|
| 1021 |
+
],
|
| 1022 |
+
"page_idx": 9
|
| 1023 |
+
},
|
| 1024 |
+
{
|
| 1025 |
+
"type": "text",
|
| 1026 |
+
"text": "Based on the above observations, we may conclude that GraphFormers are more accurate, meanwhile equally efficient and scalable as the conventional cascaded Transformer+GNNs. ",
|
| 1027 |
+
"bbox": [
|
| 1028 |
+
174,
|
| 1029 |
+
309,
|
| 1030 |
+
823,
|
| 1031 |
+
338
|
| 1032 |
+
],
|
| 1033 |
+
"page_idx": 9
|
| 1034 |
+
},
|
| 1035 |
+
{
|
| 1036 |
+
"type": "text",
|
| 1037 |
+
"text": "4.6 Online A/B Test on Bing Search ",
|
| 1038 |
+
"text_level": 1,
|
| 1039 |
+
"bbox": [
|
| 1040 |
+
176,
|
| 1041 |
+
356,
|
| 1042 |
+
431,
|
| 1043 |
+
371
|
| 1044 |
+
],
|
| 1045 |
+
"page_idx": 9
|
| 1046 |
+
},
|
| 1047 |
+
{
|
| 1048 |
+
"type": "text",
|
| 1049 |
+
"text": "GraphFormers has been deployed as one of the major ads retrieval algorithms on Bing Search, and it achieves highly competitive performance against the previous production system (the combination of a wide spectrum of semantic representation algorithms, including large-scale PLMs and cascaded PLMsGNNs). Particularly, the primary objective of Ads service is to maximize the revenue meanwhile increasing the user clicks. Therefore, the following three metrics are taken as the major performance indicators: $\\mathrm { R P M } ^ { 8 }$ (revenue per thousand impressions), CY (click yield), and $\\mathrm { C P C ^ { 9 } }$ (cost per click) . During our large-scale online $\\mathrm { A } / \\mathrm { B }$ test, GraphFormers significantly improves the overall RPM, CY, CPC by $1 . 8 7 \\%$ , $0 . 9 6 \\%$ and $0 . 9 1 \\%$ , respectively. A 11-day performance snapshot is demonstrated as Figure 3; it can be observed that in most of the time, all three metrics are significantly improved thanks to the utilization of GraphFormers (the daily performance are measured based on millions of impressions, thus having strong statistic significance). ",
|
| 1050 |
+
"bbox": [
|
| 1051 |
+
174,
|
| 1052 |
+
382,
|
| 1053 |
+
826,
|
| 1054 |
+
535
|
| 1055 |
+
],
|
| 1056 |
+
"page_idx": 9
|
| 1057 |
+
},
|
| 1058 |
+
{
|
| 1059 |
+
"type": "text",
|
| 1060 |
+
"text": "5 Conclusion ",
|
| 1061 |
+
"text_level": 1,
|
| 1062 |
+
"bbox": [
|
| 1063 |
+
174,
|
| 1064 |
+
556,
|
| 1065 |
+
299,
|
| 1066 |
+
573
|
| 1067 |
+
],
|
| 1068 |
+
"page_idx": 9
|
| 1069 |
+
},
|
| 1070 |
+
{
|
| 1071 |
+
"type": "text",
|
| 1072 |
+
"text": "In this paper, we propose a novel model architecture GraphFormers for textual graph representation. By having GNNs nested alongside each transformer layer of the pretrained language model, the underlying semantic of each textual node can be precisely captured and effectively integrated for high-quality textual graph representation. On top of the fundamental architecture, we introduce the two-stage progressive training strategy to further strengthen GraphFormers’ representation quality; we also simplify the model with the unidirectional graph aggregation, which eliminates the unnecessary computation cost. The experimental studies on three large-scale textual graph datasets verify the effectiveness of our proposed methods, where GraphFormers notably outperform the existing cascaded Transformer-GNNs methods with comparable running efficiency and scalability. ",
|
| 1073 |
+
"bbox": [
|
| 1074 |
+
174,
|
| 1075 |
+
588,
|
| 1076 |
+
825,
|
| 1077 |
+
713
|
| 1078 |
+
],
|
| 1079 |
+
"page_idx": 9
|
| 1080 |
+
},
|
| 1081 |
+
{
|
| 1082 |
+
"type": "text",
|
| 1083 |
+
"text": "6 Acknowledgement ",
|
| 1084 |
+
"text_level": 1,
|
| 1085 |
+
"bbox": [
|
| 1086 |
+
174,
|
| 1087 |
+
734,
|
| 1088 |
+
357,
|
| 1089 |
+
752
|
| 1090 |
+
],
|
| 1091 |
+
"page_idx": 9
|
| 1092 |
+
},
|
| 1093 |
+
{
|
| 1094 |
+
"type": "text",
|
| 1095 |
+
"text": "We are grateful to anonymous reviewers for their constructive comments on this work. The work was supported by grants from the National Natural Science Foundation of China (No. 62022077). ",
|
| 1096 |
+
"bbox": [
|
| 1097 |
+
173,
|
| 1098 |
+
767,
|
| 1099 |
+
823,
|
| 1100 |
+
795
|
| 1101 |
+
],
|
| 1102 |
+
"page_idx": 9
|
| 1103 |
+
},
|
| 1104 |
+
{
|
| 1105 |
+
"type": "text",
|
| 1106 |
+
"text": "References ",
|
| 1107 |
+
"text_level": 1,
|
| 1108 |
+
"bbox": [
|
| 1109 |
+
174,
|
| 1110 |
+
818,
|
| 1111 |
+
266,
|
| 1112 |
+
833
|
| 1113 |
+
],
|
| 1114 |
+
"page_idx": 9
|
| 1115 |
+
},
|
| 1116 |
+
{
|
| 1117 |
+
"type": "text",
|
| 1118 |
+
"text": "Hangbo Bao, Li Dong, Furu Wei, Wenhui Wang, Nan Yang, Xiaodong Liu, Yu Wang, Jianfeng Gao, Songhao Piao, Ming Zhou, et al. 2020. Unilmv2: Pseudo-masked language models for unified ",
|
| 1119 |
+
"bbox": [
|
| 1120 |
+
178,
|
| 1121 |
+
842,
|
| 1122 |
+
825,
|
| 1123 |
+
871
|
| 1124 |
+
],
|
| 1125 |
+
"page_idx": 9
|
| 1126 |
+
},
|
| 1127 |
+
{
|
| 1128 |
+
"type": "text",
|
| 1129 |
+
"text": "language model pre-training. In International Conference on Machine Learning, pages 642–652. \nPMLR. ",
|
| 1130 |
+
"bbox": [
|
| 1131 |
+
181,
|
| 1132 |
+
90,
|
| 1133 |
+
825,
|
| 1134 |
+
121
|
| 1135 |
+
],
|
| 1136 |
+
"page_idx": 10
|
| 1137 |
+
},
|
| 1138 |
+
{
|
| 1139 |
+
"type": "text",
|
| 1140 |
+
"text": "Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48. ",
|
| 1141 |
+
"bbox": [
|
| 1142 |
+
173,
|
| 1143 |
+
128,
|
| 1144 |
+
825,
|
| 1145 |
+
159
|
| 1146 |
+
],
|
| 1147 |
+
"page_idx": 10
|
| 1148 |
+
},
|
| 1149 |
+
{
|
| 1150 |
+
"type": "text",
|
| 1151 |
+
"text": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. ",
|
| 1152 |
+
"bbox": [
|
| 1153 |
+
174,
|
| 1154 |
+
167,
|
| 1155 |
+
821,
|
| 1156 |
+
210
|
| 1157 |
+
],
|
| 1158 |
+
"page_idx": 10
|
| 1159 |
+
},
|
| 1160 |
+
{
|
| 1161 |
+
"type": "text",
|
| 1162 |
+
"text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. ",
|
| 1163 |
+
"bbox": [
|
| 1164 |
+
171,
|
| 1165 |
+
219,
|
| 1166 |
+
823,
|
| 1167 |
+
250
|
| 1168 |
+
],
|
| 1169 |
+
"page_idx": 10
|
| 1170 |
+
},
|
| 1171 |
+
{
|
| 1172 |
+
"type": "text",
|
| 1173 |
+
"text": "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821. ",
|
| 1174 |
+
"bbox": [
|
| 1175 |
+
173,
|
| 1176 |
+
258,
|
| 1177 |
+
821,
|
| 1178 |
+
287
|
| 1179 |
+
],
|
| 1180 |
+
"page_idx": 10
|
| 1181 |
+
},
|
| 1182 |
+
{
|
| 1183 |
+
"type": "text",
|
| 1184 |
+
"text": "Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017a. Inductive representation learning on large graphs. In Advances in neural information processing systems, pages 1024–1034. ",
|
| 1185 |
+
"bbox": [
|
| 1186 |
+
174,
|
| 1187 |
+
297,
|
| 1188 |
+
823,
|
| 1189 |
+
327
|
| 1190 |
+
],
|
| 1191 |
+
"page_idx": 10
|
| 1192 |
+
},
|
| 1193 |
+
{
|
| 1194 |
+
"type": "text",
|
| 1195 |
+
"text": "William L Hamilton, Rex Ying, and Jure Leskovec. 2017b. Representation learning on graphs: Methods and applications. arXiv preprint arXiv:1709.05584. ",
|
| 1196 |
+
"bbox": [
|
| 1197 |
+
174,
|
| 1198 |
+
335,
|
| 1199 |
+
823,
|
| 1200 |
+
366
|
| 1201 |
+
],
|
| 1202 |
+
"page_idx": 10
|
| 1203 |
+
},
|
| 1204 |
+
{
|
| 1205 |
+
"type": "text",
|
| 1206 |
+
"text": "Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun. 2020. Gpt-gnn: Generative pre-training of graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1857–1867. ",
|
| 1207 |
+
"bbox": [
|
| 1208 |
+
176,
|
| 1209 |
+
375,
|
| 1210 |
+
823,
|
| 1211 |
+
419
|
| 1212 |
+
],
|
| 1213 |
+
"page_idx": 10
|
| 1214 |
+
},
|
| 1215 |
+
{
|
| 1216 |
+
"type": "text",
|
| 1217 |
+
"text": "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-˘ tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906. ",
|
| 1218 |
+
"bbox": [
|
| 1219 |
+
174,
|
| 1220 |
+
426,
|
| 1221 |
+
825,
|
| 1222 |
+
470
|
| 1223 |
+
],
|
| 1224 |
+
"page_idx": 10
|
| 1225 |
+
},
|
| 1226 |
+
{
|
| 1227 |
+
"type": "text",
|
| 1228 |
+
"text": "Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907. ",
|
| 1229 |
+
"bbox": [
|
| 1230 |
+
169,
|
| 1231 |
+
479,
|
| 1232 |
+
823,
|
| 1233 |
+
508
|
| 1234 |
+
],
|
| 1235 |
+
"page_idx": 10
|
| 1236 |
+
},
|
| 1237 |
+
{
|
| 1238 |
+
"type": "text",
|
| 1239 |
+
"text": "Chaozhuo Li, Bochen Pang, Yuming Liu, Hao Sun, Zheng Liu, Xing Xie, Tianqi Yang, Yanling Cui, Liangjie Zhang, and Qi Zhang. 2021. Adsgnn: Behavior-graph augmented relevance modeling in sponsored search. arXiv preprint arXiv:2104.12080. ",
|
| 1240 |
+
"bbox": [
|
| 1241 |
+
173,
|
| 1242 |
+
518,
|
| 1243 |
+
823,
|
| 1244 |
+
561
|
| 1245 |
+
],
|
| 1246 |
+
"page_idx": 10
|
| 1247 |
+
},
|
| 1248 |
+
{
|
| 1249 |
+
"type": "text",
|
| 1250 |
+
"text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019a. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. ",
|
| 1251 |
+
"bbox": [
|
| 1252 |
+
174,
|
| 1253 |
+
570,
|
| 1254 |
+
823,
|
| 1255 |
+
614
|
| 1256 |
+
],
|
| 1257 |
+
"page_idx": 10
|
| 1258 |
+
},
|
| 1259 |
+
{
|
| 1260 |
+
"type": "text",
|
| 1261 |
+
"text": "Zhenghao Liu, Chenyan Xiong, Maosong Sun, and Zhiyuan Liu. 2019b. Fine-grained fact verification with kernel graph attention network. arXiv preprint arXiv:1910.09796. ",
|
| 1262 |
+
"bbox": [
|
| 1263 |
+
173,
|
| 1264 |
+
622,
|
| 1265 |
+
823,
|
| 1266 |
+
652
|
| 1267 |
+
],
|
| 1268 |
+
"page_idx": 10
|
| 1269 |
+
},
|
| 1270 |
+
{
|
| 1271 |
+
"type": "text",
|
| 1272 |
+
"text": "Yi Luan, Jacob Eisenstein, Kristina Toutanova, and Michael Collins. 2020. Sparse, dense, and attentional representations for text retrieval. arXiv preprint arXiv:2005.00181. ",
|
| 1273 |
+
"bbox": [
|
| 1274 |
+
174,
|
| 1275 |
+
661,
|
| 1276 |
+
823,
|
| 1277 |
+
691
|
| 1278 |
+
],
|
| 1279 |
+
"page_idx": 10
|
| 1280 |
+
},
|
| 1281 |
+
{
|
| 1282 |
+
"type": "text",
|
| 1283 |
+
"text": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Distributed representations of words and phrases and their compositionality. arXiv preprint arXiv:1310.4546. ",
|
| 1284 |
+
"bbox": [
|
| 1285 |
+
173,
|
| 1286 |
+
700,
|
| 1287 |
+
825,
|
| 1288 |
+
729
|
| 1289 |
+
],
|
| 1290 |
+
"page_idx": 10
|
| 1291 |
+
},
|
| 1292 |
+
{
|
| 1293 |
+
"type": "text",
|
| 1294 |
+
"text": "Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In 2014 EMNLP, pages 1532–1543. ",
|
| 1295 |
+
"bbox": [
|
| 1296 |
+
171,
|
| 1297 |
+
738,
|
| 1298 |
+
823,
|
| 1299 |
+
768
|
| 1300 |
+
],
|
| 1301 |
+
"page_idx": 10
|
| 1302 |
+
},
|
| 1303 |
+
{
|
| 1304 |
+
"type": "text",
|
| 1305 |
+
"text": "Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. arXiv preprint arXiv:1802.05365. ",
|
| 1306 |
+
"bbox": [
|
| 1307 |
+
173,
|
| 1308 |
+
777,
|
| 1309 |
+
823,
|
| 1310 |
+
820
|
| 1311 |
+
],
|
| 1312 |
+
"page_idx": 10
|
| 1313 |
+
},
|
| 1314 |
+
{
|
| 1315 |
+
"type": "text",
|
| 1316 |
+
"text": "Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. ",
|
| 1317 |
+
"bbox": [
|
| 1318 |
+
171,
|
| 1319 |
+
829,
|
| 1320 |
+
823,
|
| 1321 |
+
859
|
| 1322 |
+
],
|
| 1323 |
+
"page_idx": 10
|
| 1324 |
+
},
|
| 1325 |
+
{
|
| 1326 |
+
"type": "text",
|
| 1327 |
+
"text": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683. ",
|
| 1328 |
+
"bbox": [
|
| 1329 |
+
174,
|
| 1330 |
+
869,
|
| 1331 |
+
825,
|
| 1332 |
+
911
|
| 1333 |
+
],
|
| 1334 |
+
"page_idx": 10
|
| 1335 |
+
},
|
| 1336 |
+
{
|
| 1337 |
+
"type": "text",
|
| 1338 |
+
"text": "Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. 2020. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–16. IEEE. ",
|
| 1339 |
+
"bbox": [
|
| 1340 |
+
176,
|
| 1341 |
+
90,
|
| 1342 |
+
823,
|
| 1343 |
+
133
|
| 1344 |
+
],
|
| 1345 |
+
"page_idx": 11
|
| 1346 |
+
},
|
| 1347 |
+
{
|
| 1348 |
+
"type": "text",
|
| 1349 |
+
"text": "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bertnetworks. arXiv preprint arXiv:1908.10084. ",
|
| 1350 |
+
"bbox": [
|
| 1351 |
+
171,
|
| 1352 |
+
142,
|
| 1353 |
+
825,
|
| 1354 |
+
171
|
| 1355 |
+
],
|
| 1356 |
+
"page_idx": 11
|
| 1357 |
+
},
|
| 1358 |
+
{
|
| 1359 |
+
"type": "text",
|
| 1360 |
+
"text": "Jianlin Su, Jiarun Cao, Weijie Liu, and Yangyiwen Ou. 2021. Whitening sentence representations for better semantics and faster retrieval. arXiv preprint arXiv:2103.15316. ",
|
| 1361 |
+
"bbox": [
|
| 1362 |
+
171,
|
| 1363 |
+
179,
|
| 1364 |
+
823,
|
| 1365 |
+
208
|
| 1366 |
+
],
|
| 1367 |
+
"page_idx": 11
|
| 1368 |
+
},
|
| 1369 |
+
{
|
| 1370 |
+
"type": "text",
|
| 1371 |
+
"text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. arXiv preprint arXiv:1706.03762. ",
|
| 1372 |
+
"bbox": [
|
| 1373 |
+
169,
|
| 1374 |
+
217,
|
| 1375 |
+
823,
|
| 1376 |
+
246
|
| 1377 |
+
],
|
| 1378 |
+
"page_idx": 11
|
| 1379 |
+
},
|
| 1380 |
+
{
|
| 1381 |
+
"type": "text",
|
| 1382 |
+
"text": "Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua ´ Bengio. 2018. Graph attention networks. International Conference on Learning Representations (ICLR). ",
|
| 1383 |
+
"bbox": [
|
| 1384 |
+
174,
|
| 1385 |
+
253,
|
| 1386 |
+
825,
|
| 1387 |
+
297
|
| 1388 |
+
],
|
| 1389 |
+
"page_idx": 11
|
| 1390 |
+
},
|
| 1391 |
+
{
|
| 1392 |
+
"type": "text",
|
| 1393 |
+
"text": "Chenguang Wang, Yangqiu Song, Haoran Li, Ming Zhang, and Jiawei Han. 2016a. Text classification with heterogeneous information network kernels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30. ",
|
| 1394 |
+
"bbox": [
|
| 1395 |
+
174,
|
| 1396 |
+
305,
|
| 1397 |
+
823,
|
| 1398 |
+
348
|
| 1399 |
+
],
|
| 1400 |
+
"page_idx": 11
|
| 1401 |
+
},
|
| 1402 |
+
{
|
| 1403 |
+
"type": "text",
|
| 1404 |
+
"text": "Jizhe Wang, Pipei Huang, Huan Zhao, Zhibo Zhang, Binqiang Zhao, and Dik Lun Lee. 2018. Billionscale commodity embedding for e-commerce recommendation in alibaba. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 839–848. ",
|
| 1405 |
+
"bbox": [
|
| 1406 |
+
173,
|
| 1407 |
+
356,
|
| 1408 |
+
826,
|
| 1409 |
+
412
|
| 1410 |
+
],
|
| 1411 |
+
"page_idx": 11
|
| 1412 |
+
},
|
| 1413 |
+
{
|
| 1414 |
+
"type": "text",
|
| 1415 |
+
"text": "Suhang Wang, Jiliang Tang, Charu Aggarwal, and Huan Liu. 2016b. Linked document embedding for classification. In Proceedings of the 25th ACM international on conference on information and knowledge management, pages 115–124. ",
|
| 1416 |
+
"bbox": [
|
| 1417 |
+
174,
|
| 1418 |
+
421,
|
| 1419 |
+
825,
|
| 1420 |
+
464
|
| 1421 |
+
],
|
| 1422 |
+
"page_idx": 11
|
| 1423 |
+
},
|
| 1424 |
+
{
|
| 1425 |
+
"type": "text",
|
| 1426 |
+
"text": "Wenlin Wang, Chenyang Tao, Zhe Gan, Guoyin Wang, Liqun Chen, Xinyuan Zhang, Ruiyi Zhang, Qian Yang, Ricardo Henao, and Lawrence Carin. 2019a. Improving textual network learning with variational homophilic embeddings. arXiv preprint arXiv:1909.13456. ",
|
| 1427 |
+
"bbox": [
|
| 1428 |
+
174,
|
| 1429 |
+
472,
|
| 1430 |
+
823,
|
| 1431 |
+
516
|
| 1432 |
+
],
|
| 1433 |
+
"page_idx": 11
|
| 1434 |
+
},
|
| 1435 |
+
{
|
| 1436 |
+
"type": "text",
|
| 1437 |
+
"text": "Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhiyuan Liu, Juanzi Li, and Jian Tang. 2019b. Kepler: A unified model for knowledge embedding and pre-trained language representation. arXiv preprint arXiv:1911.06136. ",
|
| 1438 |
+
"bbox": [
|
| 1439 |
+
173,
|
| 1440 |
+
523,
|
| 1441 |
+
826,
|
| 1442 |
+
566
|
| 1443 |
+
],
|
| 1444 |
+
"page_idx": 11
|
| 1445 |
+
},
|
| 1446 |
+
{
|
| 1447 |
+
"type": "text",
|
| 1448 |
+
"text": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144. ",
|
| 1449 |
+
"bbox": [
|
| 1450 |
+
173,
|
| 1451 |
+
575,
|
| 1452 |
+
825,
|
| 1453 |
+
632
|
| 1454 |
+
],
|
| 1455 |
+
"page_idx": 11
|
| 1456 |
+
},
|
| 1457 |
+
{
|
| 1458 |
+
"type": "text",
|
| 1459 |
+
"text": "Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826. ",
|
| 1460 |
+
"bbox": [
|
| 1461 |
+
171,
|
| 1462 |
+
640,
|
| 1463 |
+
823,
|
| 1464 |
+
670
|
| 1465 |
+
],
|
| 1466 |
+
"page_idx": 11
|
| 1467 |
+
},
|
| 1468 |
+
{
|
| 1469 |
+
"type": "text",
|
| 1470 |
+
"text": "Zenan Xu, Qinliang Su, Xiaojun Quan, and Weijia Zhang. 2019. A deep neural information fusion architecture for textual network embeddings. arXiv preprint arXiv:1908.11057. ",
|
| 1471 |
+
"bbox": [
|
| 1472 |
+
171,
|
| 1473 |
+
678,
|
| 1474 |
+
823,
|
| 1475 |
+
707
|
| 1476 |
+
],
|
| 1477 |
+
"page_idx": 11
|
| 1478 |
+
},
|
| 1479 |
+
{
|
| 1480 |
+
"type": "text",
|
| 1481 |
+
"text": "Cheng Yang, Zhiyuan Liu, Deli Zhao, Maosong Sun, and Edward Y Chang. 2015. Network representation learning with rich text information. In IJCAI, volume 2015, pages 2111–2117. ",
|
| 1482 |
+
"bbox": [
|
| 1483 |
+
171,
|
| 1484 |
+
714,
|
| 1485 |
+
823,
|
| 1486 |
+
744
|
| 1487 |
+
],
|
| 1488 |
+
"page_idx": 11
|
| 1489 |
+
},
|
| 1490 |
+
{
|
| 1491 |
+
"type": "text",
|
| 1492 |
+
"text": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint arXiv:1906.08237. ",
|
| 1493 |
+
"bbox": [
|
| 1494 |
+
173,
|
| 1495 |
+
752,
|
| 1496 |
+
825,
|
| 1497 |
+
795
|
| 1498 |
+
],
|
| 1499 |
+
"page_idx": 11
|
| 1500 |
+
},
|
| 1501 |
+
{
|
| 1502 |
+
"type": "text",
|
| 1503 |
+
"text": "Michihiro Yasunaga, Rui Zhang, Kshitijh Meelu, Ayush Pareek, Krishnan Srinivasan, and Dragomir Radev. 2017. Graph-based neural multi-document summarization. arXiv preprint arXiv:1706.06681. ",
|
| 1504 |
+
"bbox": [
|
| 1505 |
+
173,
|
| 1506 |
+
803,
|
| 1507 |
+
825,
|
| 1508 |
+
847
|
| 1509 |
+
],
|
| 1510 |
+
"page_idx": 11
|
| 1511 |
+
},
|
| 1512 |
+
{
|
| 1513 |
+
"type": "text",
|
| 1514 |
+
"text": "Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 974–983. ",
|
| 1515 |
+
"bbox": [
|
| 1516 |
+
173,
|
| 1517 |
+
854,
|
| 1518 |
+
826,
|
| 1519 |
+
911
|
| 1520 |
+
],
|
| 1521 |
+
"page_idx": 11
|
| 1522 |
+
},
|
| 1523 |
+
{
|
| 1524 |
+
"type": "text",
|
| 1525 |
+
"text": "Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2020. Graph neural networks: A review of methods and applications. AI Open, 1:57–81. \nJie Zhou, Xu Han, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2019. Gear: Graph-based evidence aggregating and reasoning for fact verification. arXiv preprint arXiv:1908.01843. \nJason Zhu, Yanling Cui, Yuming Liu, Hao Sun, Xue Li, Markus Pelger, Liangjie Zhang, Tianqi Yan, Ruofei Zhang, and Huasha Zhao. 2021. Textgnn: Improving text encoder via graph neural network in sponsored search. arXiv preprint arXiv:2101.06323. ",
|
| 1526 |
+
"bbox": [
|
| 1527 |
+
171,
|
| 1528 |
+
90,
|
| 1529 |
+
826,
|
| 1530 |
+
236
|
| 1531 |
+
],
|
| 1532 |
+
"page_idx": 12
|
| 1533 |
+
}
|
| 1534 |
+
]
|
parse/train/yILzFBjR0Y/yILzFBjR0Y_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/yILzFBjR0Y/yILzFBjR0Y_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|