On the Convergence to a Global Solution of Shuffling-Type Gradient Algorithms

Stochastic gradient descent (SGD) algorithm is the method of choice in many machine learning tasks thanks to its scalability and efficiency in dealing with large-scale problems. In this paper, we focus on the shuffling version of SGD which matches the mainstream practical heuristics. We show the convergence to a global solution of shuffling SGD for a class of non-convex functions under over-parameterized settings. Our analysis employs more relaxed non-convex assumptions than previous literature. Nevertheless, we maintain the desired computational complexity as shuffling SGD has achieved in the general convex setting.

Paper

Similar papers

Peer review

Reviewer JrBJ5/10 · confidence 4/52023-06-28

Summary

This paper provides the convergence to a global solution of `shuffling' SGD (where the index set is permuted and apply gradient updates according to the permutation order) for a class of non-convex functions, in the over-parameterized setting.

Strengths

The paper provides a good comprehensive background of previous work on the convergence of shuffling SGD. The theoretical results are stated clearly, and experiments on real data sets are implemented.

Weaknesses

The big weakness of the paper is that the class of non-convex functions under consideration is not defined using interpretable or standard assumptions. In particular, I am skeptical that this non-convex class is weaker than previously considered function classes (in particular, the PL condition of Nguyen et al.) The problem is that the quantities in Assumption 5 are not calculable beyond the convex setting; only in the convex setting is it clear what the constant N in Assumption 5 is (where N=0). Beyond the convex setting, there is no way to measure N and in turn, gauge the sample complexity results presented in this paper.

Questions

Please provide bounds on the constants in Assumption 5 for some concrete example classes of non-convex functions which were not included in previous global convergence proofs.

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

2 fair

Contribution

2 fair

Limitations

yes

Authorsrebuttal2023-08-12

Follow up on the rebuttal

Dear Reviewer JrBJ, We hope our responses answer all your concerns. Please let us know if you have some time considering them. In case you need any remaining clarifications, we would be more than happy to discuss with you (within this discussion period). If your concerns are all properly addressed, we really hope that the reviewer positively re-evaluates our work. Regards, Authors

Reviewer JrBJ2023-08-13

Thank you for the detailed response.

Authorsrebuttal2023-08-15

Follow up

Dear Reviewer JrBJ, Since the Author-Reviewer discussion phase will end in a fews days, we would like to follow up and discuss with you. Please do not hesitate to contact us if there are additional answers or explanations that we can make to convince you of the significance of our paper within this discussion period. We appreciate your timely response, as it would provide us with an opportunity to address any remaining questions. We emphasize that our main contributions are in the form of a new developed theoretical framework to show the convergence to a global solution of shuffling type gradient algorithms. We have also provided some experiments to confirm our theory. Our theoretical results are significant since we are the first to obtain the convergence to a global solution for this class of non-convex problems. To the best of our knowledge, we are the first to provide the ‘average PL’ condition and show that it holds for a wide class of neural networks. Therefore, our theoretical results are significant: finding a global solution for a class of non-convex problems is an important and non-trivial problem. If your concerns are all properly addressed, we really hope that the reviewer positively re-evaluates our work to support this research direction. We appreciate your inputs and we thank you for your time spent reviewing this paper. Best regards, Authors

Reviewer MeiV6/10 · confidence 3/52023-07-05

Summary

In this paper, the authors demonstrate the convergence of the shuffling version of Stochastic Gradient Descent (SGD) to a global solution for a class of non-convex functions. Notably, the analysis of the algorithm relies on more relaxed non-convex assumptions compared to previous works, while maintaining the same computational complexity as shuffling SGD in the convex setting.

Strengths

1. The paper is clear and well written. Moreover, the motivation for the paper is well-explained. 2. The authors assume more relaxed assumptions compared to previous work in the non-convex setting. The authors explain carefully the assumption and the contribution upon previous works. 3. By leveraging these assumptions, the authors derive a novel complexity bound for achieving convergence to a global minimizer. Notably, this complexity matches the performance of unified shuffling methods in previous works, which aimed for convergence to a stationary point.

Weaknesses

A potential weakness of the paper is that the analysis, upon closer inspection, appears to follow a relatively straightforward path and bears similarities to previous works. In this context, the main distinguishing factor is the introduction of the average PL condition, which specifically enables the establishment of a bound for the shuffling version of SGD.

Questions

In regards to Assumption 3, the authors claim that it is a weaker condition compared to assuming the PL inequality for every component function. However, when examining the example provided in Theorem 1 to satisfy Assumption 3, it appears that the presented example actually satisfies the PL inequality for every component function. Is there are any well-known loss functions that fulfill Assumption 3 without necessarily satisfying the stronger condition of PL inequality for every component function?

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

2 fair

Limitations

yes

Authorsrebuttal2023-08-12

Follow up on the rebuttal

Dear Reviewer MeiV, We hope our responses answer all your questions! In case you need any remaining clarifications, we would be more than happy to reply. Please let us know your thoughts as soon as you can (within this discussion period). If your questions are all properly addressed, we really hope that you consider increasing your score to support our work. Regards, Authors

Reviewer MeiV2023-08-17

Thanks for the response!

Thanks to the authors for answering my questions!

Reviewer 7CVX6/10 · confidence 3/52023-07-05

Summary

Shuffling-type SGD is widely used in practice, for which the prior works only provide gradient norm guarantees in nonconvex settings. This paper studies the convergence to a global solution for shuffling-type gradient algorithms under nonconvex settings, under a new analytical framework of more relaxed nonconvex assumptions, i.e., average PL condition and generalized star-smooth-convex condition.

Strengths

1. The authors generalize two assumptions -- PL condition and star-M-smooth-convex condition -- to investigate the convergence to global solutions in nonconvex settings. Although average PL condition: cannot be directly compared with PL condition, in the interpolated settings (all component functions share the same minimizer) the average PL condition is weaker than PL condition and also covers certain neural networks with squared loss functions. 2. The authors provide a convergence guarantee of shuffling-type SGD towards global solutions under new assumptions, while the prior works only investigate either the global convergence for SGD sampling without replacement or the gradient norm guarantees for shuffling-type SGD.

Weaknesses

1. I appreciate the discussion on the comparison between the average PL condition and PL condition the authors made in the main paper, and the applicability of average PL condition in neural networks. However, I am not sure about the comparison between these assumptions beyond the interpolated settings (which seems a strong and restricted assumption). 2. For the generalized star-M-smooth-convex condition (Assumption 5), I understand that the term $\frac{1}{n}\sum_{i = 1}^{n}||w_i^{(t)} - w_0^{(t)}||^2$ allows for more flexibility as the inner product term on the R.H.S. can be negative. But it still looks a bit unnatural to me, especially it depends on the all the iterates from the next (future) cycle generated by the algorithm. 3. Here is my major concern. The final result obtained in Theorem 2 with **increasing** learning rates only guarantees the convergence to global solutions within a fixed neighborhood of radius $O(\hat{\epsilon})$. The increasing learning rate does not match the empirical practice.

Questions

1. The notation in Section 4 is confusing. The authors used $\epsilon$ in Assumption 4, but in Line 254 they say ``$\hat{\epsilon}$ in Assumption 4''. I feel that the authors are meant to use $\hat{\epsilon}$ for Assumption 4. 2. Could the authors comment on the comparison between the average PL condition and PL condition when $\sigma_*^2 \neq 0$? 2. Could the authors explain the reason of adding the term $\frac{1}{n}\sum_{i = 1}^{n}||w_i^{(t)} - w_0^{(t)}||^2$ in Assumption 5? Does it help the analysis? Can this term be changed to only dependent on $w_{i - 1}^{(t)}$? Because from the discussion in the paper, this term is added to allow negative $\langle \nabla f(w;i) - \nabla f(w_*; i), w - w_* \rangle$. 3. In prior works for shuffling-type SGD, the coefficient in front of $\sigma_*^2$ in the convergence bound like Eq. (17) in Theorem 2 usually depends on the stepsize $\eta_t$, then one can take $\eta_t = O(1/T)$ to guarantee exact convergence to the solutions. However, I see the constant coefficient for $\sigma_*^2$ in Eq. (17), and the authors are choosing increasing stepsizes as $\eta_t = K\eta_{t - 1}$ where $K > 1$ (meaning the initial stepsize needs to be very small). Could the authors explain the intuition for this stepsize choice and the reason not getting stepsize-dependent coefficient?

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

Please see the weaknesses and questions above.

Authorsrebuttal2023-08-12

Follow up on the rebuttal

Dear Reviewer 7CVX, We hope our responses answer all your concerns. Please let us know if you have some time considering them. In case you need any remaining clarifications, we would be more than happy to discuss with you (within this discussion period). If your concerns are all properly addressed, we really hope that the reviewer positively re-evaluates our work. Regards, Authors

Reviewer 7CVX2023-08-13

I thank the authors for their detailed responses, which address my questions. I am increasing my score to 6.

Reviewer VXbB6/10 · confidence 3/52023-07-07

Summary

The paper analyzes convergence properties of shuffling-type methods (shuffle SGD and incremental gradient) to the global solution under relaxed assumptions for non-convex overparameterized functions. Particularly, the paper relaxes frequently used PL-condition to the averaged PL of local functions $f_i$, which is a weaker assumption under interpolation regime, and also imposes a weak star-smooth convex condition.

Strengths

- The paper is well written and easy to follow. All the assumptions are well explained in the paper. - Related works are well explained. - I did not check the proof details but the convergence rate looks correct (although it seems it is possible to improve convergence rate for shuffle SGD variants).

Weaknesses

1. My main concern is that the set of assumptions is not well motivated: would be nice to see examples of functions that do not satisfy PL-condition, but satisfy your relaxed conditions. 2. It is unclear how tight this bound is under proposed assumptions. Assuming PL, it is possible to achieve faster $\epsilon^{-1/2}$ convergence. Is it not possible to achieve under your set of assumptions ? 3. The paper advertises a lot analyzing shuffling-type algorithms, however its convergence rates are same as for the incremental gradient methods. For random reshuffling and single shuffling SGD faster dependence on the number of functions $n$ is possible to achieve. 4. I did not understand why in experiments the initial labels are changed. Wouldn’t it be enough just to take a large over-parameterized model ? Modifying the training data seems strange to me, especially because usually the purpose of the theory is to explain the realistic training scenarios. Coming up with new datasets that explain the theoretical results makes it look like the theory doesn’t apply to realistic training scenarios.

Questions

1. Is assuming both Assumption 5 and Assumption 3 still weaker than just assuming PL (Def 1) ? 2. Is there a relation between a constant N and sigma ? (to compare the rates in Table 1). 3. Is the epsilon in line 241 the same as the epsilon in Assumption 4 ? If not, change the notation. 4. I did not find the definition of parameter $\lambda$ (used in line 241) in the paper. 5. Which variant of shuffling SGD was used in experiments ?

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

4 excellent

Contribution

3 good

Limitations

yes

Authorsrebuttal2023-08-12

Follow up on the rebuttal

Dear Reviewer VXbB, We hope our responses answer all your questions! In case you need any remaining clarifications, we would be more than happy to reply. Please let us know your thoughts as soon as you can (within this discussion period). If your questions are all properly addressed, we really hope that you consider increasing your score to support our work. Regards, Authors

Reviewer VXbB2023-08-16

I would like to thank the authors for their detailed replies to my questions. My main remaining concern is in justification of the class of functions considered. First, I have hoped for a more formal discussion why Example 1 is not PL. Second, since assuming Assumptions 3 & 5 is neither stronger, nor weaker than PL, but leads to a worse convergence rate, it is unclear why one should use Assumptions 3 & 5. Giving some important example that is covered by new assumptions but not covered by PL might solve this concern.

Authorsrebuttal2023-08-18

Reply to Reviewer VXbB

Dear Reviewer VXbB, Thank you for your insightful question! We are happy to discuss this in detail. Please see our answers below: > First, I have hoped for a more formal discussion why Example 1 is not PL. We can show this by a simple example (satisfying conditions in Theorem 1) as follows. Let $\{( {x}^{(i)},y^{(i)})\}_{i=1}^2$ be a training data set where $ {x}^{(i)} \in \mathbb{R}$ is the input data and $y^{(i)} \in \mathbb{R}$ is the output data for $i = 1, 2$. Let us choose $z(\theta; i) = w_2 {x}^{(i)}$, so $h(w;i) = w_1 (w_2 {x}^{(i)}) + b$, where $w = [ w_1 ; w_2 ; b ] \in \mathbb{R}^3$. (This is a small two-layer neural network). We have $f(w ; i) = \frac{1}{2} || h(w;i) - y^{(i)} ||^2 = \frac{1}{2} ( w_1 (w_2 {x}^{(i)}) + b - y^{(i)} )^2$ for $i = 1, 2$. Note that $F(w) = \frac{1}{n} \sum_{i=1}^n f(w ; i)$ and $\nabla F(w) = \frac{1}{n} \sum_{i=1}^n \nabla f(w ; i)$. We have $\nabla f (w; i) = [ w_2 {x}^{(i)} ( w_1 w_2 {x}^{(i)} + b - y^{(i)} ) ; w_1 {x}^{(i)} ( w_1 w_2 {x}^{(i)} + b - y^{(i)} ) ; w_1 w_2 {x}^{(i)} + b - y^{(i)} ]$ To demonstrate a specific case, we consider the data $x^{(1)} = 1$, $x^{(2)} = 0$, $y^{(1)} = 1$, $y^{(2)} = 2$. (Other data may work as well). We observe that, $w_* = [ 1 ; -1 ; 2 ]$ is the optimal solution of $F$ since $F_* = \frac{1}{2} \sum_{i=1}^2 \frac{1}{2} ( w_1^* (w_2^* {x}^{(i)}) + b^* - y^{(i)} )^2 = \frac{1}{2} \cdot \frac{1}{2} [(1 \cdot (-1) \cdot 1 + 2 - 1)^2 + (1 \cdot (-1) \cdot 0 + 2 - 2)^2] = 0$. On the other hand, there exists a stationary point $\hat{w} = [ 0 ; 0 ; 3/2 ]$ such that $\nabla F(\hat{w}) = \frac{1}{2} \sum_{i=1}^2 \nabla f(\hat{w} ; i) = \frac{1}{2} ( [ 0 \cdot 1 ( 0 \cdot 0 \cdot 1 + 3/2 - 1 ) ; 0 \cdot 1 ( 0 \cdot 0 \cdot 1 + 3/2 - 1 ) ; 0 \cdot 0 \cdot 1 + 3/2 - 1 ] + [ 0 \cdot 0 ( 0 \cdot 0 \cdot 0 + 3/2 - 2 ) ; 0 \cdot 0 ( 0 \cdot 0 \cdot 0 + 3/2 - 2 ) ; 0 \cdot 0 \cdot 0 + 3/2 - 2 ] ) = [ 0 ; 0 ; 0 ]$ while $F(\hat{w}) = \frac{1}{2} \cdot \frac{1}{2} [ (0 \cdot 0 \cdot 1 + 3/2 - 1)^2 + (0 \cdot 0 \cdot 0 + 3/2 - 2)^2 ] = 0.125 \neq F_*$. Therefore, this example does not satisfy the PL condition on $F$ since it does not satisfy $|| \nabla F(\hat{w})||^2 \geq 2 \mu [F({\hat{w}})-F_*]$ for $\mu > 0$. Note that this is just an example to demonstrate that the general network in Theorem 1 does not necessarily satisfy the PL condition on $F$, but it satisfies our proposed Assumption 3. > Second, since assuming Assumptions 3 & 5 is neither stronger, nor weaker than PL, but leads to a worse convergence rate, it is unclear why one should use Assumptions 3 & 5. Giving some important example that is covered by new assumptions but not covered by PL might solve this concern. An example that helps to demonstrate our theory is the class of convex functions (but non-strongly convex). Note that the non-strongly convexity on $F$ cannot imply the PL condition on $F$. The example in Theorem 1 contains a class of non-strongly convex problem (e.g. $z (\theta ; i ) = x^{(i)}$), so the Assumptions 3 and 5 hold. While our theory obtains the same rate of convergences as the class of non-strongly convex problems, it does not restrict the problem to be convex. Thus, we believe our theoretical setting will be one of the directions to investigate non-convex problems with special structures. We hope our answers address all of your questions. Again, we are grateful for your positive assessment of our paper. We hope you find this direction is interesting and support us. Sincerely, Authors

Reviewer VXbB2023-08-18

I would like to thank the authors for their detailed response. I have no further questions, and I raise my score to 6.

Authorsrebuttal2023-08-21

Thank you!

Dear Area Chair and Reviewers, We would like to thank the AC and the reviewers for your time and effort in handling and reviewing our paper. Best regards, Authors

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC