Aiming towards the minimizers: fast convergence of SGD for overparametrized problems

Modern machine learning paradigms, such as deep learning, occur in or close to the interpolation regime, wherein the number of model parameters is much larger than the number of data samples. In this work, we propose a regularity condition within the interpolation regime which endows the stochastic gradient method with the same worst-case iteration complexity as the deterministic gradient method, while using only a single sampled gradient (or a minibatch) in each iteration. In contrast, all existing guarantees require the stochastic gradient method to take small steps, thereby resulting in a much slower linear rate of convergence. Finally, we demonstrate that our condition holds when training sufficiently wide feedforward neural networks with a linear output layer.

Paper

References (39)

Scroll for more · 27 remaining

Similar papers

Peer review

Reviewer APCi6/10 · confidence 3/52023-07-06

Summary

This work proposes a new condition called the aiming condition, which looks similar to quasar-convexity but provides fundamentally different convergence guarantees for SGD. Under the aiming condition, along with several other regularity conditions, SGD can achieve the same sample complexity as GD. It is then shown that wide neural networks enjoys this property with high probability.

Strengths

- The aiming condition is impressive as a condition for SGD to achieve the same sample complexity as GD. - The presentation of the results is clear. This work embodies a lot of results, which could potentially make the paper hard to follow. However, the introduction part provides a clear roadmap.

Weaknesses

- I am expecting more discussions on the comparison of the aiming condition against existing conditions (e.g. quasar-convexity).This involves two aspects: 1. I would like to see some more examples where the aiming condition is satisfies but quasar-convexity does not hold; 2. It would be interesting to provide some intuition about why quasar-convexity cannot provide a similar result. - This work does not seem to be enough strong technically. - This work has some minor problems presentations. For example, the same sentence appears twice in Line 62 and Line 160; Furthermore, in Line 251, the radius needs to be scaled up rather than shrunk.

Questions

I am mainly curious about the comparison of quasar-convexity and the aiming condition (see "Weaknesses" above).

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

This work, with a theorical nature, does not have potential negative societal impact.

Reviewer vP9H6/10 · confidence 3/52023-07-06

Summary

This work shows a regularization condition for SGD in the interpolation regime which allows it to have same fast linear convergence rate as deterministic gradient descent. Hence, the theory presented in this paper supports the practical observation that with the same (large) learning rate, mini-batch SGD has almost the same convergence rate of SGD. This goes against the traditional approaches for SGD convergence under PL condition that require SGD to have smaller step-size than GD, hence making SGD's convergence slower in theory. The authors here present the conditions in which SGD have similar iteration complexity as GD.

Strengths

1) The main strength of the work lies in it's novelty. The motivation of the problem seems clear from the introduction. 2) This is an important problem since it is crucial to have a large learning rate for SGD as it helps improving generalization. Hence, the current theories that analyze SGD under PL-condition and requires it to have a smaller l.r, may suffer from having good generalization. The theory presented by the authors allows the same large learning rate to be used as GD. Hence, the variance of SGD coupled with the alowable large learning rate can boost generalization.

Weaknesses

1) Looking at the theorems, I still feel the assumptions made are too strong too hold in a non-convex landscape especially Aiming. For theorem-1.2, which consdiers a non-convex loss landscape on $w$, it is unclear how aiming can hold on a ball of radius $r$ and it's relation to the curvature at that point (given by minimal eigenvalue of NTK). The authors make sure that the iterates remain inside a ball $B_{r}(w_{0})$ near the minima but the radius depends on the eigenvalue of NTK at initializaiton! This makes a very strong assumption that the initial point is set to be very close to the true minima $w_{0}$. More so, if initialization is made far away from the minima, there is a high probability that aiming does not hold. 2) From the theorems, it is unclear that how the stochasticity from the mini-batch gradients effect the convergence or what is it's effect in ensuring that the iterates don't escpae $B_{r}(w_{0})$. For theorem-2.3, I assume that it will take longer iterations if the variance in mini-batch gradient is high. It is not very clear from the inequality as how the relation is. A minor correction: 3) On line-42, it might be better to cite a different source such as [1] to refer to the generalization effect of large learning rate, as the authors of "Edge of stability" themselves mention that they don't claim the generalization effect of large learning rate but rather focus on the stability effect. See https://youtu.be/6xeh6gfESuc?t=3016 [1] Li, Yuanzhi, Colin Wei, and Tengyu Ma. "Towards explaining the regularization effect of initial large learning rate in training neural networks." Advances in Neural Information Processing Systems 32 (2019).

Questions

See points 1 and 2.

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The main limitation of this work is a strong assumption on the locality of the SGD analysis. Some further clarification would be helpful.

Reviewer FvUq6/10 · confidence 3/52023-07-07

Summary

This work presents a set of conditions under which the convergence rate of SGD with large step size is similar to that of gradient descent (the deterministic setting). This is in contrast to prior work, where the convergence of over-parameterized SGD under PL condition require small step size, and converge slowly as a result of this small step size.

Strengths

1. The paper is well-motivated and clearly organized. This work provides a theoretical result that improve previous analyses of the linear convergence rate of overparameterized SGD. 2. The paper thoroughly compares its result to prior works mentioned. 3. The paper attempts verifies assumptions that it makes.

Weaknesses

1. The paper states that its goal is to improve stepsize selection and convergence rate of SGD for nonconvex problems under certain conditions (see lines 133-134). However, the paper does not directly do this. 2. This work claims to be the first to present an analysis of the convergence rate of over-parameterized SGD with large step size which is similar the convergence rate attained by gradient descent,. However, prior work touches upon this as well (see “Observations in simplified settings” in The Impact of Neural Network Overparameterization on Gradient Confusion and Stochastic Gradient Descent by Sankararaman, De, Xu, Huang, and Goldstein). Perhaps it is worth connecting both works to further demonstrate this paper’s contribution. (Please correct me if I am mistaken.) 3. While the contribution of this paper is predominantly theoretical, I believe that this work could benefit from some empirical evaluation as well. Current experiments are fairly limited (MNIST is the only dataset used and is easy to learn. Additionally, model architecture is also quite limited. Using other datasets e.g. CIFAR10/100 or running the experiments for networks with varying widths might be useful).

Questions

There is no figure 2

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

The authors have described the limitations.

Reviewer eCMo5/10 · confidence 2/52023-07-09

Summary

This paper studies the convergence of SGD with large step size. It is shown that under some regularity conditions, SGD enjoys a fast linear convergence rate, both in expectation and with high probability. These results can be applied to show fast convergence of SGD for wide enough feed-forward networks.

Strengths

The paper is well organized and easy to follow, and especially the authors have compared in detail the assumptions/results with those in the existing papers. It is indeed an important question to study the convergence of SGD with large step sizes, and the results in this paper seem novel and interesting.

Weaknesses

1. The main application of the results is for wide enough neural networks in the NTK regime, which seems restrictive. 2. It would be helpful if the authors can highlight the technical novelties. 3. It is discussed around line 40-42 that large batch sizes could be beneficial for generalization. However, the current results don't seem to have such implications. In particular, I don't think the regime studied in the current paper is a case of the "edge of stability" regime. 4. Is it possible to verify the proposed conditions for common neural network models, at least empirically? Otherwise it's not clear if the developed theory reflects what happens in practice.

Questions

Please see the question above.

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

2 fair

Limitations

The authors have adequately addressed the limitations.

Reviewer APCi2023-08-17

Thanks for the rebuttal

I would like to thank the authors for the detailed rebuttal. I certainly had some misunderstandings when writing the initial review, and I appreciate the authors for the great explanation. I will vote for this work being accepted.

Reviewer eCMo2023-08-18

Response to the authors

I thank the authors for the detailed response. I don't have further questions.

Reviewer vP9H2023-08-19

i acknowledge that I have read the author's response and they have answered my concerns. I intend to keep my score.

Reviewer FvUq2023-08-21

I acknowledge and appreciate the authors' responses. I intend to keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC