SLaM: Student-Label Mixing for Distillation with Unlabeled Examples

Knowledge distillation with unlabeled examples is a powerful training paradigm for generating compact and lightweight student models in applications where the amount of labeled data is limited but one has access to a large pool of unlabeled data. In this setting, a large teacher model generates ``soft'' pseudo-labels for the unlabeled dataset which are then used for training the student model. Despite its success in a wide variety of applications, a shortcoming of this approach is that the teacher's pseudo-labels are often noisy, leading to impaired student performance. In this paper, we present a principled method for knowledge distillation with unlabeled examples that we call Student-Label Mixing (SLaM) and we show that it consistently improves over prior approaches by evaluating it on several standard benchmarks. Finally, we show that SLaM comes with theoretical guarantees; along the way we give an algorithm improving the best-known sample complexity for learning halfspaces with margin under random classification noise, and provide the first convergence analysis for so-called ``forward loss-adjustment" methods.

Paper

Similar papers

Peer review

Reviewer 7BQA5/10 · confidence 4/52023-06-21

Summary

This paper aims to systematically utilize the noise of teacher predictions in a knowledge distillation setting comprised of both labeled and unlabeled samples. The proposed method, SLaM linearly mixes the student predictions with the noise of the teacher predictions, thereby hoping that the student will learn the noise-free true underlying label. The method is supported by both theoretical results and empirical experiments on different datasets.

Strengths

- The paper is well-written, easy to follow, and nicely structured. - The proposed method, SLaM is relatively simple and general which is a nice feature. - The method is supported by both theoretical and empirical results.

Weaknesses

I am generally positive about the paper, but my biggest concern is the empirical evaluation of the methods. I am open to changing my score if the following is addressed. Comments and questions: - All empirical results presented are based on "[...] the best test accuracy over all epochs.". I.e. the methods are continuously evaluated on the test set (effectively making it a validation set), and no real test set performance is available - this is bad practice. Thus, I am concerned whether the reported scores are actually representative of the real test performance of the methods, and thereby if one can, in fact, rely on the conclusions drawn. I would strongly suggest reporting something like the test performance at the last epoch to avoid overfitting the test set. And even so, the experiments might be biased towards having evaluated the methods on the test set previously, so ideally evaluations on completely independent test sets would be included. - In e.g. L230 SLaM is reported to "[...] consistently outperform the baselines, often by a large margin". I believe this is too bold a claim. In particular, only 7 of 24 experiments yield improvements over $1\%$ compared to the second-best baseline, while 8 other of 24 have improvements $<0.2\%$ of which 3 have improvements $<0.02\%$. Furthermore, there are 2 of 24 experiments where the method does not improve over the baselines. All of these observations are without consideration of the variance, which likely makes some additional improvements statistically indistinguishable. I would suggest reducing the claims slightly, especially given the evaluation protocol. - It is unclear why this method shouldn't be considered a semi-supervised learning method, as both labeled and unlabeled samples are provided and used for the distillation procedure. In fact, the availability of a labeled validations set, $V$ appears to be a necessity for this method, and I believe it should be presented as a semi-supervised setting more explicitly. - A lot of results and details are deferred to the appendix. While I fully understand the limitations a 9-page restriction can put on a paper, I believe some of the space in the paper could be better utilized to present at least some key results from the appendix. Minor comments: - L148: "agree" -> "disagree" - L166: "etting" -> "setting" - L311: "inFigure 2" -> "in Figure 2" - Figure 1 is never referenced in the paper - Figure 2 is badly formatted; i.e. too small a font and hard to distinguish the results/lines of different methods (consider rescaling the y-axis to e.g. improvement over baseline)

Questions

See the above.

Rating

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

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

Sufficiently addressed.

Reviewer Un325/10 · confidence 4/52023-07-02

Summary

The paper introduces SLaM, a method for knowledge distillation using unlabeled examples. Its main contribution lies in the development of an efficient and data-agnostic approach that enhances student model performance in scenarios with limited labeled data but abundant unlabeled data. SLaM accomplishes this by incorporating information from both labeled and unlabeled examples during the distillation process. The paper extensively presents experimental results showcasing SLaM's superiority over previous methods on standard benchmarks. Furthermore, the authors provide theoretical guarantees and insights into SLaM's effectiveness. Overall, the paper delivers a valuable contribution to the field of knowledge distillation, showcasing a novel approach supported by empirical evidence.

Strengths

1. The paper introduces a novel approach to knowledge distillation with unlabeled examples that leverages the information from both labeled and unlabeled examples during the distillation process. 2. The experimental results show the effectiveness of the proposed method in knowledge distillation. 3. Theoretical results further guarantee the effectiveness of the proposed method.

Weaknesses

The proposed method argues that existing methods suffer from the shortcoming of noisy pseudo-label, but it also seems hard to guarantee the proposed method can get the pseudo-labels without noisy. Moreover, the proposed method can be limited to semi-supervised learning problem.

Questions

1. Can the proposed method ensure that the pseudo-labels generated by SLaM are noise-free, thereby addressing the limitation discussed in the paper? How and why? 2. Although not specifically designed for knowledge distillation, the baseline method Taylor-CE shows competitive or even superior performance on celebA even with only 2% labeled data. Can you elaborate on the reasoning behind this, especially highlighting the advantages of the proposed method in dealing with sparsely labeled data, which Taylor-CE lacks? 3. It is better to plot a figure to demonstrate the performance of the comparison method in different percents of the labeled training data with a large span. Eg: 5%, 10%, 30%, 50%, 70%. 4. The baseline method Taylor-CE, which focuses on noise label problem, also achieve competitive or even superior performance. I wonder if the proposed method can work in noise label learning that also need reliable pseudo-labels? For example, on clothing1m dataset.

Rating

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

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

Please refer to weakness and question.

Reviewer PQzZ7/10 · confidence 3/52023-07-02

Summary

In their paper, the authors propose a knowledge distillation procedure that augments pseudo-labels predicted by a teacher models by its uncertainty, leading to a uncertainty-aware student learning routine. This is realized by a novel loss formulation that mixes student predictions with quantified uncertainty associated with the pseudo-labeling prediction. Together with convincing empirical results demonstrating improvements in terms of generalization performance, theoretical results characterize the convergence of the proposed solution.

Strengths

- Strong empirical results against recent baselines - Profound theoretical results - Well-written manuscript with a clear and valid motivation, addressing a relevant problem - The simplicity of the approach is appealing, promoting the applicability of this method

Weaknesses

Major: - While the empirical evaluation itself is certainly convincing from a generalization point of view, a more rigorous ablation study of the influence of the uncertainty quantification would be insightful. Namely, it is hard to judge how sensible this procedure is towards a good quality of the isotonic regression solution. One possibility to address this concern could be by obtaining $\alpha(x)$ and $k(x)$ from an oracle with “perfect” estimations, and gradually contaminating these estimates. Then, one could see how the performance evolves. - Relating to the previous comment, I am wondering why exactly isotonic regression is chosen as a means to quantify $\alpha(x)$. I see little explanations on why exactly this method is chosen, and not others, like conformal prediction. Of course, it is completely valid to make an arbitrary design choice that works well, but I would love to see a more thorough elaboration on the motivation of this specific construction. An ablation study as described before could be used to support this decision. Minor: - The figures are relatively small and hard to read, enlarging them would be appreciated - The orthography could be improved, many missing commas (e.g., when starting sentences like “In Section xyz, …”) - The references are inconsistent (e.g., the conference identifiers: NeurIPS, ICML, … vs. International Conference on Learning Representations)

Questions

-

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

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

I do not see too much about limitations, this could certainly be deepened.

Reviewer E5Wk6/10 · confidence 3/52023-07-08

Summary

The paper attempts to resolve the problem of noisy teacher labels for unlabeled examples in knowledge distillation. Proposed Student-Label Mixing (SLaM) to introduce noise into student prediction when matching to teacher’s pseudo-label on unlabeled data. Experiments on standard benchmarks were used to demonstrate the effectiveness of the proposed method. Also, the paper provided theoretical guarantees for SLaM for learning noisy halfspaces. rating updated after author rebuttal.

Strengths

The paper studied the critical problem about how to learn unlabeled examples for the student model with knowledge distillation when the teacher model is not perfect. The paper proposed to resolve it from a novel angle by introducing noise to the student model’s prediction, and also provided a theoretical guarantee for the proposed solution. Extensive experiments based on common datasets and multiple baselines were provided to demonstrate its effectiveness.

Weaknesses

The paper lacks the explanation on the intuition of the mixing solution on formula (1), where there is no clarification why it is essential for the second term, i.e. the element-wise multiplication term. Even though the paper developed a solution to consider the noise for student model prediction, it seems not robust enough to learn \alpha(x) and k(x). As it requires an additional validation dataset, it is easy to make \alpha overfit to those data. Also it has another strong assumption that \alpha is ​​isotonic regression. All of these make the learning of \alpha easily overfit the data which cannot be robust enough, which also make the solution not that elegant. It lacks the result details about \alpha and k setup for each experiment. Also it lacks the related ablation study to demonstrate how \alpha and k affect the model performance. no comparison with a baseline where the validation dataset is used for vanilla distillation, and on top that add vanilla unlabeled distillation or other state of the art unlabeled distillation.

Questions

The proposed algorithm has the additional cost of needing a validation dataset, seems a fair baseline would be to incorporate the validation set for vanilla distillation, and on top that add vanilla unlabeled distillation or other state of the art unlabeled distillation? What is the intuition of the second term on formula (1)? How important is the design \alpha(x) and k(x) is? How does the size of the validation set affect the overall performance?

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

yes

Reviewer 7BQA2023-08-11

- Changing to last-epoch evaluations would be an improvement, but I still believe it should be followed by some comments on the improper evaluation scheme. In particular, since I assume the last-epoch evaluation is still based on particular successful experiments, based on the test performance (which should not really be possible beforehand). - The suggested change of phrasing would be adequate for me, however, I would also suggest at least modifying the use of "consistently" as a few experiments are not improved and a few more have negligible improvements (<0.02). - My main concern is that both in the title and the main paper, a lot of emphases is put on the wording "distillation with unlabeled examples", but since the method does require a labeled validation dataset, I believe the naming is misleading.

Authorsrebuttal2023-08-13

Once again we thank the Reviewer for their comments and feedback. > Changing to last-epoch evaluations would be an improvement, but I still believe it should be followed by some comments on the improper evaluation scheme. In particular, since I assume the last-epoch evaluation is still based on particular successful experiments, based on the test performance (which should not really be possible beforehand). We do understand the Reviewer's concern —  after all, this is a well-known issue that has been raised within the ML community throughout the years. (Unfortunately, the fact is that papers that use standard academic datasets like CIFAR and Imagenet — including every baseline we compared against — do use this evaluation scheme and so we had to follow the same approach to guarantee a fair comparison.) To address the Reviewer's concern: (i) We will add some comments on the evaluation scheme we used that reflect the Reviewer's concern and our discussion above. (We are happy to adopt any suggestions regarding the phrasing of the latter.) (ii) We will include the ablation-experiments which show that our approach is robust to parameters like the size of the validation datasets, the choice of the regression method for learning a(x) and k(x), errors in the estimates of a(x) and k(x) (please see the general response for more details). Hopefully this mitigates the danger of overfitting to the nuances of the academic datasets chosen for evaluation. > The suggested change of phrasing would be adequate for me, however, I would also suggest at least modifying the use of "consistently" as a few experiments are not improved and a few more have negligible improvements (<0.02). We will modify the use of "consistently" and change our phrasing to: "Our approach performs on par with previous SOTA approaches and, interestingly, it often outperforms them by a large margin in the important cases where (i) only a few labeled examples are available; (ii) one deals with large-scale problems with many classes like the Imagenet dataset". We are happy to adopt any further suggestions from the reviewers. > My main concern is that both in the title and the main paper, a lot of emphases is put on the wording "distillation with unlabeled examples", but since the method does require a labeled validation dataset, I believe the naming is misleading. The term "distillation with unlabeled examples" is not ours — it was introduced in [CKSNH] to describe the very popular distillation-setting where one has access to only a few labeled examples but a lot of unlabeled examples. Note also that using a few of the labeled examples as a validation set is a very common practice in these type of "learning with few labeled examples"-settings and, additionally, as we mention in Lines 274-278 of the main body of the paper, to be fair to methods not using validation data, we have included the validation data in the training dataset of all methods. Another term that has been used to describe the setting we consider is "semi-supervised distillation", but we find it a bit confusing since in standard semi-supervised learning one does not have access to a teacher model. Does the reviewer prefer the latter term/have another suggestion? [CKSNH]: Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, Geoffrey Hinton. Big Self-Supervised Models are Strong Semi-Supervised Learners. NeurIPS 2020.

Reviewer 7BQA2023-08-14

I appreciate the reply and effort from the authors toward all reviewers. The proposed changes do improve the contribution and remove some concerns for me. However, I will keep my current score and suggest for accept. **"Another term that has been used to describe the setting we consider is "semi-supervised distillation", but we find it a bit confusing since in standard semi-supervised learning one does not have access to a teacher model. Does the reviewer prefer the latter term/have another suggestion?"** I understand the AI/ML field has an abundance of unclear and duplicate terminology as well as methods with two different namings. However, I believe the inclusion of distillation in "semi-supervised distillation" rightfully states the use of both labeled and unlabeled data as well as a teacher model, whereas "distillation with unlabeled examples" lacks the labeled data reference. Thus, while I personally find "semi-supervised distillation" more accurate, I am not to force this change on you. However, I think this discussion warrants a mention in the paper to avoid any misconceptions about the use of labeled/unlabeled data.

Authorsrebuttal2023-08-15

We would like to thank the reviewer for their suggestion. We will include a discussion on the terms semi-supervised distillation and distillation with unlabeled examples.

Reviewer PQzZ2023-08-13

Thanks for the author's response, and the efforts spent into further complementing the insights. The additions improve the overall contribution, which makes me even more confident in proposing to accept this paper.

Authorsrebuttal2023-08-13

We greatly appreciate the Reviewer's feedback and effort in evaluating our work. We will adopt all of their suggestions and add remarks that summarize our discussion.

Reviewer Un322023-08-16

Thanks for the authors’ response. **To A1**, the author said the pseudo-labels produced by the proposed method can also be noisy, but the author motivates the new method due to existing methods suffering from the shortcoming of noisy pseudo-label. It seems the proposed method does not address the shortcoming of the existing method. Moreover, from the comparison result below, we can see that the benefits can be marginal, which demonstrates the proposed method is not substantially different from the comparison method. **To A2**, the author does not answer my question Q2. SLaM beats Taylor-CE in other experiments cannot explain why the baseline method Taylor-CE beats the proposed method.

Authorsrebuttal2023-08-16

Clarifications needed

>To A1, the author said the pseudo-labels produced by the proposed method can also be noisy, but the author motivates the new method due to existing methods suffering from the shortcoming of noisy pseudo-label. It seems the proposed method does not address the shortcoming of the existing method. We show that student models trained with SLaM will produce **better** pseudo-labels than other methods. **It is impossible to generate perfect pseudo-labels and no method achieves that**. Perhaps the Reviewer could clarify what they mean by “suffering from the shortcoming of noisy pseudo-label” and “It seems the proposed method does not address the shortcoming of the existing method.” >To A2, the author does not answer my question Q2. SLaM beats Taylor-CE in other experiments cannot explain why the baseline method Taylor-CE beats the proposed method. In our previous response we tried to explain that, since Taylor-CE only beats our method in 1 out of the 28 experiments we considered, we feel that there is no trend to be explained here — that particular data-point is most probably a random outlier. In general, the advantage of our method compared to Taylor CE (and any other generic label-noise method) is that our method is designed to provably exploit the structure of the instance specific noise from the teacher model — this is why it outperforms Taylor CE in 27 out of the 28 experiments, and by a large margin in the most difficult dataset, i.e., Imagenet. We feel that we may have not understood the Reviewer’s question (because, frankly, we see no trend justifying that Taylor CE in particular is better than our method), so maybe the Reviewer could further clarify what they are asking? >Moreover, from the comparison result below, we can see that the benefits can be marginal, which demonstrates the proposed method is not substantially different from the comparison method. We respectfully disagree with the reviewer: the improvements are **not marginal** as there are several cases where SLaM outperforms baselines by large margins (sparsely labeled data cases), while all methods converge to the same performance as the noise from the teacher is decreasing (as expected). See the paragraph “Performance Gains of SLaM as a Function of The Number of Labeled Examples “ for more details.

Reviewer E5Wk2023-08-17

thanks for addressing the questions

thanks for addressing the questions. I increased my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC