Shuffling Gradient-Based Methods for Nonconvex-Concave Minimax Optimization

This paper aims at developing novel shuffling gradient-based methods for tackling two classes of minimax problems: nonconvex-linear and nonconvex-strongly concave settings. The first algorithm addresses the nonconvex-linear minimax model and achieves the state-of-the-art oracle complexity typically observed in nonconvex optimization. It also employs a new shuffling estimator for the"hyper-gradient", departing from standard shuffling techniques in optimization. The second method consists of two variants: semi-shuffling and full-shuffling schemes. These variants tackle the nonconvex-strongly concave minimax setting. We establish their oracle complexity bounds under standard assumptions, which, to our best knowledge, are the best-known for this specific setting. Numerical examples demonstrate the performance of our algorithms and compare them with two other methods. Our results show that the new methods achieve comparable performance with SGD, supporting the potential of incorporating shuffling strategies into minimax algorithms.

Paper

References (44)

Scroll for more · 32 remaining

Similar papers

Peer review

Reviewer QEyv7/10 · confidence 3/52024-06-20

Summary

This work propose shuffling gradient-based methods for nonconvex-linear and nonconvex-strongly convex minimax optimization and obtain complexity results.

Strengths

The presentation is clear. Based on my knowledge on both shuffling and minimax optimization, I found that the algorithms and complexity results are reasonable. The experiments provide sufficient details for reproducibility.

Weaknesses

The contribution is in general combinational, with slightly more techniques such as smoothing and shuffling-based function evaluations. Also, comparison with previous stochastic algorithms for minimax optimization and shuffling for nonconvex optimization is lacking, as shown in question 1 below.

Questions

(1) You could compare your complexity results (Theorem 2) with previous stochastic gradient-based minimax optimization works (e.g. [a]) and shuffling for nonconvex optimization (e.g. [b]). [a] Lin, T., Jin, C., \& Jordan, M. (2020, November). On gradient descent ascent for nonconvex-concave minimax problems. In International Conference on Machine Learning (pp. 6083-6093). PMLR. [b] Nguyen, L. M., Tran-Dinh, Q., Phan, D. T., Nguyen, P. H., \& Van Dijk, M. (2021). A unified convergence analysis for shuffling-type gradient methods. Journal of Machine Learning Research, 22(207), 1-44. (2) At the bottom of page 1, for problem (NL), it looks possible to remove matrix $K$ and use $F_i:\mathbb{R}^p\to \mathbb{R}^q$, since $\langle F_i(w),Ku\rangle=\langle K^{\top}F_i(w),u\rangle$ and we could replace $K^{\top}F_i(w)\in \mathbb{R}^q$ with $F_i(w)\in \mathbb{R}^q$. (3) In Assumption 2, do you mean for any $v\in {\rm dom}(h)$, $||v||_2\le M_h$? (4) In "Condition (6) can be relaxed to the form $\frac{1}{n} \sum_{i=1}^n||\nabla F_i(w)-\nabla F(w)||^2 \leq \sigma_J^2+\Theta_J||\nabla \Phi_0(w)||^2$", should $\Phi_0$ be $F$? (5) After Assumption 5, ``If $f=0$, then $\mathcal{G} _ {\eta}(w)\equiv\nabla\Phi_{\gamma}(w)$''. Should it be $\nabla\Phi_0(w)$ instead of $\nabla\Phi_{\gamma}(w)$, since the definition of $\mathcal{G} _ {\eta}(w)$ in eq. (18) involves $\nabla\Phi_0(w)$ not $\nabla\Phi _ {\gamma}(w)$? Or you might consider defining $\mathcal{G} _ {\eta,\gamma}(w)$ with $\nabla\Phi_0(w)$ replaced by $\nabla\Phi _ {\gamma}(w)$, such that $\epsilon$-stationary point of (3) and that of (10) in Lemma 2 correspond to $\mathcal{G} _ {\eta,0}(w)$ and $\mathcal{G} _ {\eta,\gamma}(w)$ respectively. (6) I found that Eq. (21) (Option 1) needs to save {$F_{\pi^{(t)}(j)(w_0^{(t)})}$}$_{j=1}^n$ to avoid repeated evaluation of these $n$ values. Yes? (7) Line 255: "fix number" $\rightarrow$ "fixed number". (8) At the end of Theorem 2, do you mean to change one of the $\nabla_w\mathcal{H}_i$ into $\nabla_u\mathcal{H}_i$? It seems that the complexity of semi-shuffling is bettter than full-shuffling. Then what's the advantage of full-shuffling?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The checklist mentions the limitation that this work only focuses on nonconvex-linear and nonconvex-strongly convex minimax optimization, and says "We do not yet know if our paper has an immediate broader impact. However, since our problems and our algorithms are sufficiently general, we hope they will create broader impacts." I agree with both of them.

Reviewer jEDb5/10 · confidence 3/52024-07-12

Summary

This paper focuses on nonconvex-concave, finite-sum (stochastic) minimax problems with possibly nonsmooth regularization. Aiming to find $\epsilon$-stationary points, the paper proposes shuffling-based proximal gradient descent-ascent algorithms and verifies gradient computation complexity upper bounds via a bilevel optimization perspective. The results consider various settings, such as (1) the nonconvex-linear (NL) case with either strongly convex or convex $h$, each requiring $\mathcal{O}(n \epsilon^{-3})$ and $\mathcal{O}(n \epsilon^{-7/2})$ gradient evaluations, respectively, and (2) the nonconvex-strongly-concave (NS) case with smoothness assumptions, where semi-shuffling requires $\mathcal{O}(n \epsilon^{-3})$ while full-shuffling requires (naively) $\mathcal{O}(n \epsilon^{-4})$ gradient evaluations. The paper contains proofs of the convergence theorems and supporting numerical experiments.

Strengths

* The theoretical results and the problem-specific proof techniques could be a solid contribution. To the best of my knowledge, this is one of the very few works that discuss shuffling-type algorithms for minimax problems apart from Das et al., 2022 [1] and Cho & Yun, 2023 [2]. I think the results under the relaxed setting, especially the linear, non-strongly-convex case of (NL), and for the bilevel-type shuffling algorithms are quite novel. The paper also includes experimental results that align with the theoretical results.

Weaknesses

* Despite that the novelty of this work would be in the fact that the results apply for broader settings than previous literature rather than obtaining *faster* convergence under similar settings, the paper still should have included a fair, more detailed comparison with previous work and illustrated how the gradient evaluation complexity becomes different from settings with stronger assumptions. The current draft does not contain any meaningful quantitative comparison with previous work. * In particular, I don’t think that the settings in [2] are different from the results for the (NS) setting. Apart from the fact that the objective function does not contain $f$ and $h$, the assumptions used in Theorem 1 of [1] (Assumptions 1-4) and the NS setting seem to be nearly identical, and the goal of finding $\epsilon$-stationary points also seem to be equivalent. (In fact, for the variable $u$ the PŁ assumption is weaker than strong concavity.) In terms of the convergence rates, Theorem 1 of [1] requires only $\mathcal{O} (n \epsilon^{-2} + n^{1/2} \epsilon^{-3})$ gradient evaluations, which is smaller than $\mathcal{O} (n \epsilon^{-3})$. The authors should have at least explained how introducing $f, h$ changes the difficulty of the problem, or given any other reasons why one could prefer the bilevel optimization framework over the simple SGDA algorithm as in [1].

Questions

- Why is it necessary to use bilevel optimization, i.e., update $n$ steps of $w$ and then $n$ steps of $v$, instead of iteration-wise updates? Also, am I right that the results for the (NS) settings are weaker than that of [2] at least when $f, h = 0$? Is there a particular reason why the weaker convergence rates are inevitable if we consider $f, h$? - Some technical parts of the algorithms are brought from previous work, such as the use of $\Phi_{\gamma}$ inspired by Yang et al., 2020 [3]. Some other parts look new, such as the placement of the proximal steps considering $f$ and $h$ and the variants of algorithms like choosing between (21), (22) in Algorithm 1 and (26), (27) in Algorithm 2. I am curious about the motivations of the new components of the algorithms, especially considering the different types of shuffling techniques I mentioned last—whether these have clear motivations or are more like artifacts of the proof techniques, if there were similar approaches in minimization problems, etc. - Light question: Would it be able to relax the strongly concave case to weaker PŁ-type assumptions, or consider alternating type updates as in [2]? Moreover, is it possible to compare the presented results for minimax problems with nonconvex, nonsmooth *minimization* as in Cutkosky et al., 2023 [4]? - Light question: Are there examples of when the nonconvex-linear (NL) setting could be of interest in practice? - Minor: Why are there two bibliographies, and why are the contents different? It seems that the one at the end of the appendix has the correct numbers.

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

There seems to be no potential negative societal impact.

Reviewer czC85/10 · confidence 2/52024-07-12

Summary

This paper proposes shuffling gradient-based methods for addressing two classes of minimax optimization problems: nonconvex-linear and nonconvex-strongly concave settings. The first algorithm focuses on the nonconvex-linear minimax setting and the second algorithm, consisting of semi-shuffling and full-shuffling schemes, focuses on the nonconvex-strongly concave minimax setting. The authors establish oracle complexity bounds for these algorithms under some standard assumptions. Numerical experiments are conducted.

Strengths

The authors establish oracle complexity bounds for their algorithms under standard assumptions. These bounds provide theoretical guarantees on the performance of the proposed methods.

Weaknesses

The proposed Algorithm 1 requires solving a maximization problem at each iteration, which can be computationally expensive. The numerical experiment section is quite weak. First, only Algorithm 1 is tested, with no results related to Algorithm 2. Second, only one minimax model is tested. More complex and popular minimax models should be included, and state-of-the-art competitors should be used to evaluate the performance of the proposed methods. Typo in Line 187 and Line 459: $M_f \rightarrow M_f^2$

Questions

What are the advantages of the proposed methods in this paper, given the introduction of the semi-shuffling and full-shuffling schemes? Do they have a better convergence rate or sample complexity compared to existing methods?

Rating

5

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

yes

Reviewer 4w6E6/10 · confidence 3/52024-07-13

Summary

The paper presents new shuffling gradient-based methods for solving two classes of nonconvex-concave minimax optimization problems: nonconvex-linear and nonconvex-strongly concave settings. The first algorithm is designed for the nonconvex-linear setting and achieves state-of-the-art oracle complexity, employing a new shuffling estimator for the hyper-gradient. The second method introduces semi-shuffling and full-shuffling schemes for the nonconvex-strongly concave setting, establishing their oracle complexity bounds for the first time. Numerical examples are provided to demonstrate the performance of the proposed algorithms.

Strengths

1. The paper is well-written, offering a clear exposition of the current state of the art, along with explicit assumptions and conditions for the proposed methods. 2. The authors introduce novel smoothing techniques similar to Nesterov's smoothing approach for lower-level maximization problems, and establish oracle complexity bounds for both nonconvex-linear and nonconvex-strongly concave settings. This introduces a novel perspective on addressing nonconvex optimization problems. 3. The paper presents two distinct algorithms, each carefully tailored to different problem settings, showcasing a profound comprehension of diverse problem landscapes.

Weaknesses

1. **Complexity Comparison with Existing Works**. The authors are encouraged to present a comparative table that delineates the oracle complexity of their proposed methods alongside that of existing works, including standard nonshuffling methods. This side-by-side comparison should aim to explicitly demonstrate the theoretical advantages of their methods. 2. **Shuffling Methods vs. Nonshuffling Methods**. The paper should discuss the distinctions between shuffling methods and traditional nonshuffling approaches. Specically, it is better to clarify whether shuffling methods consistently yield superior performance over nonshuffling methods theoretically. 3. **Numerical Experiments Expansion**. While the experiments have been conducted on synthetic datasets and small-scale datasets, which is customary for theoretical papers, the scope of empirical validation could be enhanced. It would be much better if the authors could extend their experiments to include large-scale datasets.

Questions

See above.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer QEyv2024-08-08

Reviewer QEyv is satisfied with the authors' response and will increase rating to 7

Reviewer QEyv is satisfied with the authors' response and will increase rating to 7.

Authorsrebuttal2024-08-09

Thank you for your support!

Dear Reviewer QEyv, We are glad that you are satisfied with our response and increase the rating score. This is very encouraging! Thank you very much for your support! Best regards, Authors

Authorsrebuttal2024-08-09

Follow up on the rebuttal

Dear Reviewer czC8, 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

Authorsrebuttal2024-08-11

Follow up

Dear Reviewer czC8, Since the Author-Reviewer discussion phase will end in two 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 clarify our paper within this discussion period. We appreciate your timely response, as it would provide us with an opportunity to address any remaining questions. If your concerns are all properly addressed, we really hope that the reviewer positively re-evaluates our work. We appreciate your inputs and we thank you for your time spent reviewing this paper. Best regards, Authors

Reviewer czC82024-08-13

Thank you for the detailed response to my questions. I will raise my score accordingly.

Authorsrebuttal2024-08-09

Follow up on the rebuttal

Dear Reviewer jEDb, 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

Authorsrebuttal2024-08-14

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 Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC