Implicit Contrastive Representation Learning with Guided Stop-gradient

In self-supervised representation learning, Siamese networks are a natural architecture for learning transformation-invariance by bringing representations of positive pairs closer together. But it is prone to collapse into a degenerate solution. To address the issue, in contrastive learning, a contrastive loss is used to prevent collapse by moving representations of negative pairs away from each other. But it is known that algorithms with negative sampling are not robust to a reduction in the number of negative samples. So, on the other hand, there are algorithms that do not use negative pairs. Many positive-only algorithms adopt asymmetric network architecture consisting of source and target encoders as a key factor in coping with collapse. By exploiting the asymmetric architecture, we introduce a methodology to implicitly incorporate the idea of contrastive learning. As its implementation, we present a novel method guided stop-gradient. We apply our method to benchmark algorithms SimSiam and BYOL and show that our method stabilizes training and boosts performance. We also show that the algorithms with our method work well with small batch sizes and do not collapse even when there is no predictor. The code is available at https://github.com/bych-lee/gsg.

Paper

Similar papers

Peer review

Reviewer CW7J6/10 · confidence 4/52023-06-27

Summary

The paper proposes the implicit contrastive learning algorithm, which uses the guided stop-gradient to push away negative samples without the uniformity term in the contrastive loss. By applying the method to the non-contrastive methods including SimSiam and BYOL, the paradigm combines the advantages of contrastive and non-contrastive algorithms and boosts the downstream performance in various downstream tasks.

Strengths

1. The improvements brought by GSG are significant and the experiments are solid. 2. The main idea and insights of GSG are easy to follow. 3. The combination between the asymmetric architectures and contrastive loss is interesting.

Weaknesses

1. Besides the guided stop-gradient, simply combing the asymmetric architectures and the contrastive loss (e.g., InfoNCE) seems would have the same effect. So what is the advantage of implicit contrastive loss? Is it possible that the advantages of GSG with small batch sizes are brought by the asymmetric architecture? More comparisons between the implicit contrastive loss and the explicit contrastive loss would make this paper more solid. 2. In Section 6.2, the authors show that GSG contributes to training stability by removing the predictors. However, the final accuracy is still far from the results with the predictor. Is it possible to show that the GSG stabilizes the training process of SimSiam with the predictor? It would be better to show more differences between SimSiam and SimSiam with GSG during the training process. 3. When selecting where to apply stop-gradient, the paper uses the distance in the projection layer. However, the loss is calculated in the prediction layer. Besides the brief explanations on page 4, it would be better to provide more theoretical or empirical evidence.

Questions

See Weaknesses

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer imWm6/10 · confidence 4/52023-07-03

Summary

This article presents a way to improve the learning of non-contrastive self-supervised learning methods such as BYOL and SimSiam. It uses incorporates implicitly contrastive notions, by removing elements of the loss that may lead to close representations collapsing together. This modification is proven to improve the representations for BYOL and SimSiam with different experiments. In particular, variations of the algorithm with other Stop Gradients are shown to be detrimental to learning, and the utility of the algorithm to prevent collapse at low batch sizes or without projectors is shown.

Strengths

The presented method is simple but provides consistent improvement for both BYOL and the SimSiam learning methods. In particular, the method improves low batch sizes as well as when the predictor is removed. The method is ablated to consider different Stop Gradients variations. The article is clear and well written.

Weaknesses

I find the article a bit lacking in ablations and inquiries about the method. For instance: - I would have liked a variation of the algorithm with more than only 2 examples. Do using N examples for the implicit contrast improves further the method? - And the method prevents collapse without a predictor, but the improvement is more unclear in the general case. Some measure of dimensionality for instance, or something similar, would have helped demonstrate the improvement of the method is due to helping a potential collapse of the representations. (I do not find the t-SNE visualisations to be convincing) I am surprised by the accuracy results presented for the benchmark of the method. Exploring Simple Siamese Representation Learning, Chen et. al, 2021 provide in Table 4 accuracies for SimCLR and Moco which are much closer to the SimSiam accuracies than the one presented in this article. In particular, the End-to-End accuracies seem very low. For these reasons, I am a bit unconvinced by the benchmark. I found the 3 Figures redundant for explaining the method. Similarly, I found the method to be a bit overexplained. The paper in its present state is a bit light, which is why I am only considering a weak accept for now.

Questions

Computing distances between representations is quite costly. I would expect a slowdown due to the computation for the method. Is it present, or is it compensated by the reduced number of computations needed by the Stop Gradients? Have other alternatives, such as reweighting the "attracting" part of the loss rather than simply removing them been considered? Do the authors have some hypotheses about the links between this work and explicit contrastive learning methods?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors have adequately addressed the limitations of the article, albeit in the appendix.

Reviewer Wcuk6/10 · confidence 4/52023-07-04

Summary

This paper proposes a novel SSL technique that can be applied on top of SimSiam or BYOL to select where to apply asymmetric predictor. This method first computes embeddings (before the predictor) and compute relevant distances. Then based on this distance, it chooses where to apply the predictor. Experiment results show improvement of this method over SimSiam/BYOL baseline.

Strengths

1. The idea is well motivated, based on the dynamics of SimSiam/BYOL representation space. 2. Experiment shows the effectiveness of the proposed method. I appreciate full evaluation including linear probe, kNN, transfer learning, and detection. 3. Ablation of several different strategies for choosing predictors is very insightful.

Weaknesses

1. Overall the innovation is limited. It is an interesting trick over SimSiam/BYOL, but not transferrable to general SSL architecture. 2. Experiment is questionable. Using batch size 512 is not considered optimal with ImageNet pre-training. In fact, the imagenet linear probe numbers reported for the BYOL baseline are lower than the original paper, with this batch size (it is supposed to be only -0.5% lower than using 4k batch size). I understand that it is only trained on 8 NVIDIA A100 GPUs. But it is possible that hyperparameters are no longer optimal.

Questions

1. Does this method effectively double the batch size? How exactly each `pairs` of images are sampled? What is memory/compute overhead?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer cNZ67/10 · confidence 4/52023-07-05

Summary

The paper introduces the Guided Stop-Gradient (GSG) method that can be applied to SSL algorithms that adopt asymmetric dual encoders such as BYOL and SimSiam in order to boost their performance and stabilize their training. The idea of the GSG is to augment the loss function to attract different views of two different images instead of just one as it is done in BYOL and SimSiam. However, the method does not explicitly repel the representations of the views of the different images but carefully selects which positive views to attract with a stop-gradient operation. Experiments with pretraining are performed on BYOL and SImSiam models using ImageNet and CIFAR10 datasets and their trained model is applied on several downstream tasks.

Strengths

- The method is simple and practical to implement. - Results are consistent and demonstrate increased performance on pretraining and downstream tasks. - Using this method, models can be trained with lower batch size which is a plus.

Weaknesses

I believe that the paper in the current form lacks analysis (and intuition) that would help understanding why the method works better than the baselines. For example, I would like to see more analysis on the structure of the learned representation space, and the effect of the choice of the batch size and number of “negative” images added to the loss. Given that the experiments are performed on smaller datasets, this is feasible to add. See also my questions below.

Questions

I have the following questions: 1. I find that choices of k in Section 5.1. unintuitive. In the case of CIFAR10, wouldn’t a large k give more insights into how well the latent space is separated? On the other hand, using k=200 for ImageNet seems very large. Could you report the variation in labels for the retrieved neighbors, i.e., is the majority vote significant? This would be interesting to analyze also in case of CIFAR10 for larger k. 2. In linear probing, you train the classification layer for the same amount of epochs as the backbone and on a much larger batch size, which sounds a lot. Is there a reason for increasing the batch size to 4096? Have you tried with just a few iterations and smaller batch size? 3. I was wondering if you could add some discussion and analysis that would help better understand why GSG works well. It would be particularly interesting to understand why GSG isn’t affected by changes in the batch size. My hypothesis is that this is because the model isn’t explicitly penalized for misplacing (potentially many) negative examples. This is partially addressed in Appendix E, however, I find tSNE visualization subjective. Do you have any numbers supporting the separation of clusters, for example, std of representations of each class? Another experiment could be to add more images to your loss (x3, x4, …), would that hurt or boost the performance? 4. I am wondering what is the dimension of the representation space? Just to be clear, in my understanding you use z in all your experiments? Minor: 5. In Section 5.2. it is not clear whether you finetune the backbone or use the frozen model. 6. When applying random stop-gradient in Section 6.1, do you randomly choose two terms out of 4 or do you randomly choose one term for each of the images? i.e, randomly choosing either first or second term and either third or fourth in Equation 2? 7. I believe that the claim in line 284 that the fluctuation of the accuracy in the beginning of the training of SimSiam is less severe when the model is trained with GSG is too strong without supporting numbers. Could you add some variance analysis? 8. Section 3 is helpful for understanding the idea of the model but it would be great if you could make it explicit that x1 and x2 are different images.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Code is provided but the paper checklist is not. Societal impacts nor limitations are not discussed.

Reviewer imWm2023-08-11

I thank the authors for their answers to my questions, notably on the benchmark and the computation cost. - For my question about using $N$ examples, I meant a generic value of examples and not only 2. I apologize for the use of $N$ as I did not mean the batch size in particular. Even using 3 examples could allow for a more complex formulation, by increasing the number of distances considered and thus the strength of the implicit contrast. - The addition of the relative variance does show an improvement of dimensionality. Some of my concerns have been answered, however I still find the article light since they are no ablations or extensions of the method. This method is still effective at low batch sizes which makes it a worthwile intermediate between contrastive and non-contrastive learning, so I am keeping my rating as a weak accept.

Authorsrebuttal2023-08-11

Response to Reviewer imWm

We thank the reviewer for the response. We hope the following addresses the reviewer's concerns. --- Our ablation study can be found in Section 6. There we provide: 1. To find out the effect of **Guided SG**, it was compared with other variants, Random SG and Reverse SG. 2. To find out how much the **predictor** contributes, it was compared with the case where the predictor was removed. 3. By comparing SimSiam and BYOL, the influence of the **momentum encoder** (only in BYOL) could be investigated. Note that the components of SimSiam w/ GSG are two encoders, one predictor, and the GSG method. In the case of BYOL w/ GSG, a momentum encoder is one of the two encoders. Therefore, we have dealt with the major components in our ablation study. --- The extension the reviewer mentioned requires bringing in other ideas to deal with new situations. For instance, if there are 3 examples, there are 2 reference examples $x_2$ and $x_3$ for one example $x_1$. When determining which of the two projections $z_{11}$ and $z_{12}$ to apply stop-gradient, the result considering the relationship with $x_2$ and the result considering the relationship with $x_3$ may be different. That is, one may be $z_{11}$, and the other may be $z_{12}$. At this time, we need another criterion to break the tie. More situations must be considered if we consider more than 3 examples simultaneously. However, introducing these additional ideas is not a straightforward extension. We believe different papers require different levels of extension. Since this paper aims to propose a simple but transformative idea, trying to add more to the current algorithm can make it overly complex and is beyond the scope of the paper.

Reviewer cNZ62023-08-14

reply to the rebuttal

I thank the authors for the thorough rebuttal. I especially appreciate the discussion and results in [G1] and [G2], and the extra results to answer my questions. Based on these comments I recommend to accept the paper. I increased my rating accordingly.

Reviewer Wcuk2023-08-18

I thank the authors for the response. All my questions are addressed. I will update my score accordingly.

Reviewer CW7J2023-08-20

reply to the rebuttal

I thank the authors for the responses. I will maintain my original rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC