Enhancing Sharpness-Aware Optimization Through Variance Suppression

Sharpness-aware minimization (SAM) has well documented merits in enhancing generalization of deep neural networks, even without sizable data augmentation. Embracing the geometry of the loss function, where neighborhoods of 'flat minima' heighten generalization ability, SAM seeks 'flat valleys' by minimizing the maximum loss caused by an adversary perturbing parameters within the neighborhood. Although critical to account for sharpness of the loss function, such an 'over-friendly adversary' can curtail the outmost level of generalization. The novel approach of this contribution fosters stabilization of adversaries through variance suppression (VaSSO) to avoid such friendliness. VaSSO's provable stability safeguards its numerical improvement over SAM in model-agnostic tasks, including image classification and machine translation. In addition, experiments confirm that VaSSO endows SAM with robustness against high levels of label noise.

Paper

References (57)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 988S4/10 · confidence 4/52023-06-27

Summary

This paper proposes a method, which applies EMA on batch gradients, to better solve the maximum problem in SAM and obtain higher test accuracy.

Strengths

1.The proposed method is easy to implement. 2.The theoretical analysis is sufficient.

Weaknesses

1.The core of this paper is to alleviate the gradient noise when solving the inner maximum problem of SAM. To achieve this goal, it proposes to implement EMA on the batch gradient and gives theoretical proof on the variance suppression. Although experiment results show the improvement of VaSSO, I am somehow confused that why the gradient noise harms the generalization performance of SAM. In the Section 4.1 of SAM[1], it claims that smaller batch size tends to yield models having better generalization ability, which conflicts the core idea of this paper. 2.Following 1, the statements on Line 47-48 and Line 50 need more support or evidence. And I think the introduction takes too much space to describe related works. In contrast, the author should give some experiment results to support the statement on Line 47-50. [1] Foret P, Kleiner A, Mobahi H, et al. Sharpness-aware minimization for efficiently improving generalization[J]. arXiv preprint arXiv:2010.01412, 2020.

Questions

no question

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

no limitation

Reviewer 65jc5/10 · confidence 4/52023-06-29

Summary

This paper proposes to improve upon sharpness-aware minimization (SAM) with variance-reduced inner perturbation steps. Theoretical analysis demonstrates that the proposed method achieves similar convergence rate with the original SAM. Empirical results on different applications (image classification and neural machine translation) demonstrate the effectiveness of proposed method. In addition, the proposed method also endows SAM with robustness against large label noise.

Strengths

- The proposed method is clear and easy to understand - Theoretical results are sound - Empirical results demonstrate the superiority of proposed method

Weaknesses

- Some claims are not fully supported - Some baseline methods are missing

Questions

- I am a bit puzzled by the claim that the proposed method can be easily integrated with computational efficient variants of SAM, e.g., (Liu et al, 2022; Zhao et al., 2022b) as well as a missing reference (Jiang et al., 2023). Compared with SAM that only needs to compute stochastic gradient, VaSSO has to keep track of d_t, as in (4a). I am not sure if the update is still possible if we only perform SAM step periodically (Liu et al, 2022) or randomly (Zhao et al., 2022b). The authors need to elaborate more on that. - If such integration is not possible, the proposed method may suffer from larger computational cost than these variants. Although it can be regarded as the cost of better final performance, the authors may still need to make this limitation clear. - Despite these computation efficient versions, the authors claim that some other existing works on SAM are also orthogonal to their work and can be easily integrated. While I do not find critical problems for these works, the authors may need to add some experiments that integrate the proposed method with these methods, and see if such integration can achieve any improvements. Nevertheless, such experiments are missing in current version. - It also surprises me that the experiments contain so few baseline methods. For example, why is Fisher SAM (Kim et al., 2022) not compared in experiments? It is confusing to compare with ASAM and not Fisher SAM. References: Weisen Jiang, Hansi Yang, Yu Zhang, James Kwok. An adaptive policy to employ sharpness-aware minimization. ICLR 2023

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Please see Weakness and Questions part.

Reviewer BP9H5/10 · confidence 3/52023-07-03

Summary

This paper proposes a variance suppression approach for SAM in order to account for the sensitivity of stochastic gradients used in SAM inner maximization. The proposed method is shown to provably reduce the MSE of gradient estimation. Some experimental results on benchmark datasets are provided.

Strengths

Paper strengths: 1. great motivation for variance suppression 2. simple algorithm modification to standard SAM 3. theoretical results for VaSSO in Theorem 2 and Corollary 1

Weaknesses

Paper weaknesses: 1. The authors claim in section 3.2 that VaSSO can boost performance of other SAM family approaches, but this is not shown in the experimental results 2. gains are marginal for larger datasets, see Table 3 3. authors provide convergence rates for Vasso in Corollary 1 but it's not clear that these are sharper than the SAM rates in Theorem 1

Questions

Questions to address in rebuttal: 1. does VaSSO translate to gains in domain generalization performance? e.g. WILDS benchmark [1] 2. can authors cite more Frank-Wolfe papers in connection to min-max problems, e.g. [2], [3] 3. how do convergence rates in Corollary 1 relate to Frank-Wolfe optimization convergence rates? convergence rates for FW algorithms and various variants have been studied in the literature [4] 4. what are the limitations of VaSSO? authors should include a discussion of limitations 5. can the authors add more SAM baselines to the list? Also, ASAM combined with VaSSO is not included, and would be good to include for comparisons I am willing to increase my score if the authors address most of these concerns. [1] Koh et al, WILDS: A Benchmark of in-the-Wild Distribution Shifts, https://arxiv.org/abs/2012.07421 [2] Tsiligkaridis et al, Understanding and Increasing Efficiency of Frank-Wolfe Adversarial Training, CVPR 2022, https://arxiv.org/abs/2012.12368 [3] Gidel et al, Frank-Wolfe Algorithms for Saddle Point Problems, https://arxiv.org/pdf/1610.07797.pdf [4] Huang et al, Accelerated Stochastic Gradient-free and Projection-free Methods, ICML 2020

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

Not discussed.

Reviewer zZon4/10 · confidence 4/52023-07-11

Summary

**I have read the author's rebuttal, see reply below** This work proposes changing SAM from perturbing the weights by the gradient of the current minibatch, with instead a moving average of the gradients across training iterations. They provide upper bounds that indicate that the moving average may better approximates the inner-maximization with respect to the loss over the entire dataset, which is more difficult for naive SAM due to the minibatch noise. Empirically, SAM with moving average gradients (which they call VaSSO) performs better on CIFAR10, CIFAR100 by around $0.1-0.5$%, and the improvements are more apparent in the presence of heavy label noise with gains up to +10% improvement with 75% label noise. Empirical experiments show that VaSSO minimizes the maximum eigenvalue of the Hessian better than SAM.

Strengths

**Significance**: The proposed algorithm VaSSO is a very simple change that achieves flatter minima and achieves better generalization than SAM. Although it may not necessarily be the case that two effects are tied in a causal manner, its effectiveness at achieving both beyond SAM may suggest that it may be a useful optimizer in practice and for further study in future works. **Quality**: The paper is clearly written and organized.

Weaknesses

**1. Lack of important literature review**: The authors motivate VaSSO by identifying problems with SAM in terms of its ability to optimize the original objective _due to the minibatch noise_. The moving average is hypothesized to be better at correctly approximating the full-batch gradient. The authors do not address the fact that works that have analyzed SAM previously have unanimously observed that SAM actually **only observes improvements in generalization if it is paired with minibatch noise**. In particular, Andruschenko et al (https://arxiv.org/abs/2206.06232) showed that n-SAM which utilizes the full-batch gradient directly for the perturbation step actually observed no generalization gain. This is supported by experiments on m-sharpness in the original SAM paper (https://arxiv.org/abs/2010.01412), and they also showed that better approximating the sharpest direction by also taking the second order approximation of the loss instead of the first order approximation lead to worse performance. Wen et al (https://arxiv.org/abs/2211.05729) showed that the minibatch noise is important to minimize the trace of the Hessian instead of the max eigenvalue (though the work does not make any claim about which measure of sharpness is better correlated with generalization). To summarize, the authors pose the problem as that SAM cannot achieve the full generalization benefits because it does not minimize its intended objective optimally, but previous literature indicate that this suboptimality is actually what allows SAM to achieve better generalization. The authors do not mention the conclusions made in previous works, but it seems important to address this conflict. **2. Marginal improvements, and inconsistent numbers in comparison to Foret et al.**: Without label noise VaSSO's improvements range between 0.2-0.5% for CIFAR10 and CIFAR100, and it's not clear whether further hyperparameter tuning could close this gap. In particular, the authors train WideResNet-28-10 on CIFAR10 with cutout data augmentation which was also conducted by Foret et al, where they report an error of 2.3% while the authors report 2.7% and for VaSSO 2.5%. This slight improvement may potentially come from the authors only optimizing the rho hyperparameter, and not m-sharpness (the other hyperparameter mentioned in Foret et al. where the minibatch is sharded first). For label noise, there is a much more significant boost in test accuracy, but it is unclear whether the authors are reporting the peak early-stopping accuracy or the final accuracy. The former is more important for label noise, and the difference between VaSSO and SAM may be coming from reporting the latter. Also, Foret et al reported better numbers for SAM on CIFAR10 with similar amounts of label noise although they were using ResNet34 instead of ResNet18. **3. The upper bounds are too loose, and comparing upper bounds does not lead to any meaningful comparison.** The upper bound derived in Theorem 2 requires very approximate intermediate steps (very loose upper bound in Line 181), and is not a function of $t$ but $T$. The bound implies that if you train for longer, the bound for the MSE between the moving average and the full batch gradient _at every intermediate step_ improves which doesn't seem right. More importantly, any meaningful comparison between the MSE for SAM and MSE for VaSSO should compare the upper bound of VaSSO to the _lower bound_ for SAM. Comparing two upper bounds isn't very meaningful.

Questions

See weaknesses

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

N/A

Reviewer BP9H2023-08-14

I thank the authors for their responses. I have decided to increase my score.

Reviewer 65jc2023-08-18

I have checked your responses as well as comments from other reviewers. I think the additional experiments addressed most of my concerns, so I increased my score. The authors should add these additional experiments to the camera ready version.

Authorsrebuttal2023-08-20

We are glad to address your concerns. The discussions above will be included in the polished paper for sure.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC