Cluster-aware Semi-supervised Learning: Relational Knowledge Distillation Provably Learns Clustering

Despite the empirical success and practical significance of (relational) knowledge distillation that matches (the relations of) features between teacher and student models, the corresponding theoretical interpretations remain limited for various knowledge distillation paradigms. In this work, we take an initial step toward a theoretical understanding of relational knowledge distillation (RKD), with a focus on semi-supervised classification problems. We start by casting RKD as spectral clustering on a population-induced graph unveiled by a teacher model. Via a notion of clustering error that quantifies the discrepancy between the predicted and ground truth clusterings, we illustrate that RKD over the population provably leads to low clustering error. Moreover, we provide a sample complexity bound for RKD with limited unlabeled samples. For semi-supervised learning, we further demonstrate the label efficiency of RKD through a general framework of cluster-aware semi-supervised learning that assumes low clustering errors. Finally, by unifying data augmentation consistency regularization into this cluster-aware framework, we show that despite the common effect of learning accurate clusterings, RKD facilitates a "global" perspective through spectral clustering, whereas consistency regularization focuses on a "local" perspective via expansion.

Paper

Similar papers

Peer review

Reviewer T9nK6/10 · confidence 3/52023-07-01

Summary

This paper studies relational knowledge distillation (RKD) on semi-supervised learning setting. This is different from the previous theoretical results on knowledge distillation, which mostly focus on feature matching (not relational). They introduce a new notion called low clustering error, which quantifies the difference betweeen the predicted and ground truth clusterings, and show that RKD can provably produce low clustering error. They also have a few other results, including the label efficiency theorm on a cluster-aware semi-supervised learning framework with low clustering errors, and unifying data augmentation consistency regularization with RKD, showing that consistency reuglarization focuses on the local perspective, while RKD improves over the global perspective.

Strengths

Originality: Since we already have seen a few papers on connecting spectral clustering with contrastive learning or other learning methods, the spectral clustering part of this paper is not particularly surprising. However, I do like the relational part, which focuses on the relationships between the features, rather than the feature matching itself. I think this is a very interesting and important direction to pursue. Moreover, the authors also investigate the benefit of clustering awareness in the semi-supervised learning scenario, which is novel. Quality: I think this is a well written paper with rigorous definitions, theorems and proofs. Clarity: Since this is a theory paper, I will not say it is straightforward to read. However, I think the authors did a very good job in the introduction to demonstrate the main contributions of this paper, and present the technical theorems and assumptions in a structured way. Therefore I would say this paper is well written. Significance: I think this paper is indeed interesting, because it provides an interesting connection between relational knowledge distillation learning and spectral clustering, and also provide theoretical characterization on various aspects.

Weaknesses

It is a bit difficult for me to understand the intuition of Assumption 4.2, and the toy example after assumption 4.2 is too special, which does not explain much about the complicated definitions. It would be nice to further elaborate on the intuitions, and perhaps give a non-trivial example of this assumption, and why this assumption is a mild assumption. Personally I like Theorem 4.1 better, and I do not feel it is really necessary to consider the limited unlabeled samples setting (because usually we may assume the unlabeled samples are cheap). Therefore, it might improve the presentation of this paper, if the author can add more motivations on Sec 4.2. (if there is no such motivation, it's also fine, I think it is a minor spot.) The current connections between different results are weak. Specifically, I cannot easily understand how the last DAC result connects with the previous two. I might be wrong, but I initially felt that the third result says the combination of DAC and RKD can improve the performance. However, the statement of Theorem 6.2 seems only contains DAC, not RKD. Therefore, it seems to me that the first and third results are in parallel: both of them are providing upper bounds of clustering error. Is it true? -- if this is true, I think the presentation of the introduction should be updated, to avoid potential misconceptions.

Questions

I do not have other questions.

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Not applicable.

Reviewer APrr5/10 · confidence 3/52023-07-06

Summary

This paper points out that RKD over a population-induced graph given by a teacher model leads to low clustering error (Spectral Clustering), which provably benefits to semi-supervised learning and other downstream tasks. Consequently, it provides a theoretical analysis of relational knowledge distillation (RKD) in the semi-supervised classification setting from a clustering perspective.

Strengths

This paper mainly focuses on expounding the close relationship between RKD and Spectral Clustering from a theoretic perspective, which definitely contributes to the increase of attention on RKD. According to this seminal work, RKD can be regarded as a principle choice to extract knowledge from a pre-trained large model.

Weaknesses

The weakness of this paper is obvious: The experimental part is too weak. There is no empirical study in the main text, and the experimental results in the appendix are limited. In addition, in some cases, RKD even deteriorates the Top-1 performance. The authors provided no explanation on these cases. Besides, CIFAR-10 and CIFAR-100 are not quite challenging datasets, and they are too similar. It is suggested conducting experiments on the large-scale datasets to better demonstrate the effectiveness of the proposed method.

Questions

My principle concern is the inadequate empirical study. As mentioned above, there are mainly two issues should be addressed. First, in the current results (Table 1 and Table 2), why the performance of the proposed approach cannot consistently outperform the baseline? Second, what is the performance of the proposed method on other large-scale datasets? If these issues can be resolved, I will recommend this work.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

I agree with the limitations provided in the last section. Besides, please see the Weaknesses and Questions parts for my concern. I hope the review is helpful for this work.

Reviewer mtgh6/10 · confidence 3/52023-07-08

Summary

The paper provides a comprehensive theoretical proof of the relationship between RKD and clustering from the perspective of spectral clustering. The theoretical proof presented in the paper is detailed and reasonable. Firstly, RKD is regarded as a spectral clustering problem of the teacher model, and then the introduced clustering error is used to prove that RKD leads to low clustering error. Finally, the paper presents the sampling complexity bound of RKD in the case of limited unlabeled samples.

Strengths

1.The theoretical proof presented in the paper is detailed and reasonable. 2. Label efficiency of cluster-aware SSL is demonstrated as well.

Weaknesses

1. It appears that the author attempts to convey numerous viewpoints in this paper, resulting in a somewhat confused logical structure. As a result, the paper is somewhat difficult to read and understand, possibly due to my unfamiliarity with the writing style of purely theoretical papers. 2. While the author's arguments are generally reasonable, there seems to be a lack of strong correspondence between the presented problem and the method of reasoning. It would be beneficial to understand the author's motivations and thought process behind exploring RKD at a theoretical level, particularly in relation to demonstrating its label efficiency in the SSL setting.

Questions

As I am less familiar with purely theoretical work, I find that in Section 7, the author continues to discuss the theoretical analysis of RKD in different settings. While providing relevant theoretical proofs is important based on previous empirical work, it would be beneficial to consider adding a discussion section that explores potential new directions for practical applications.

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

1. The research presented in the paper holds significant theoretical value. It would be beneficial to enhance the overall clarity and organization of the paper's structure. This will improve the readability and friendliness of the paper for fellow researchers in the field. 2. In the field of machine learning, it is highly recommended to conduct necessary experiments to strengthen the scientific rigor of the study.

Reviewer APrr2023-08-13

Please provide more explanation on the experimental results that do not quite match your claim

Thanks for offering the additional experimental results and more explanation for my questions. After reading these, most of my concerns have been resolved. However, I still have problems in understanding your rebuttal: RKD brings more improvements to FixMatch when the augmentation strength is weaker or the unlabeled sample size N is smaller, whereas when the data augmentations are strong and N is large, the gap between FixMatch and FixMatch+RKD may not be noticeable. According to your new empirical results, I noticed that this conclusion may not be true, especially when considering the **unlabeled sample size**. Besides, I found in your experimental results, the higher augmentations and more unlabeled data do not consistently lead to better results (e.g., in the Uniform&Weak case on CIFAR10 and the StochasticGreedy&50000 case on CIFAR100), which is a little strange to me. Do you have any further explanation on this issue?

Authorsrebuttal2023-08-15

Thanks for the response

Many thanks for the follow-up questions, as well as for the time and effort. Before diving into the detailed explanations for specific observations, we would like to kindly highlight the key message of our experiments: __The incorporation of RKD with the “global” perspective via spectral clustering brings improvement to DAC with the “local” perspective.__ In the updated experiments with multiple random seeds, we observe that this generally holds for different combinations of augmentation strengths and unlabeled sample sizes. Moreover, our theoretical analysis implies that the “local” perspective of DAC tends to be deficient while adding the “global” perspective of RKD can be helpful when data augmentations are not suitable (e.g., not “expansive” enough in terms of satisfying $c$-expansion property in Def 6.1 with a small $c \approx 1$) and/or when the unlabeled samples are limited. As an illustrative (but arguably oversimplified) intuition, RKD tends to bring observable improvement to DAC with weak augmentations and/or insufficient unlabeled samples. This explains why, for example in the updated experiments with $N=50000$ and high/medium augmentation strengths, the gaps between FixMatch v.s. FixMatch+RKD are not significant. __[When and how, but not how much, RKD can help]__ Although our analysis casts insight on when and how incorporating RKD brings benefits, inferring the magnitude of such improvement when scaling with the augmentation strength and unlabeled sample size is a much trickier problem in practice. In particular, for the unlabeled sample size, $N$, recall that the clustering error bounds for both RKD and DAC (Thm 4.2, 4.3, and 6.1, 6.2) depend on $N$. In other words, with insufficient unlabeled samples, the clustering accuracies of both RKD and DAC are compromised. Although their complementary perspectives suggest the advantage of RKD+DAC over DAC alone (as we observed in the experiments), such improvement may not monotonically increase as $N$ decreases. This depends on how much the clustering errors of RKD v.s. DAC are affected by decreasing $N$, which is affected not only by the algorithms themselves but also by the randomness in the training data and data augmentations. We will clarify this point in the discussion on experiments in the revision. __[Stronger augmentations/slightly more unlabeled samples do not necessarily lead to better performance]__ Despite being intuitive, we kindly point out the subtle but critical difference between “strong” and “suitable” data augmentations. Recall that in Def 6.1, data augmentations are required to be class-invariant (i.e., preserving semantic information) in addition to the expansion property. Intuitively, an overly high augmentation strength can distort the semantic information in the data and therefore hurt the performance. For the unlabeled sample size, although the accuracies of both FixMatch and FixMatch+RKD generally increase as $N$ increases, the results can fluctuate within the standard deviations (as the reviewer mentioned) since the gap between $N=40000$ and $N=50000$ in our experiments is relatively small for models like WideResNet.

Reviewer APrr2023-08-21

Thanks for the reply. My concerns are partially addressed, and I will retain my score.

Authorsrebuttal2023-08-15

Thanks for the response and feedback

We really appreciate the valuable feedback and will make sure to incorporate corresponding clarifications in our revision. In particular, we will include the following: - We will further clarify the motivation of exemplifying DAC as another instance of cluster-aware SSL with a complementary perspective to RKD in both the introduction and the associated Sec 6 (as elaborated in our response to Reviewer T9nK “__[Connection and difference between RKD and DAC]__”). - We will provide more intuitions for Ass 4.2 and further highlight the associated example concerning the pitfall of RKD alone (as elaborated in our response to Reviewer T9nK “__[Intuition of Ass 4.2]__”). - We will update the experiments and the corresponding discussion in Appendix A (as presented in our general response).

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC