Single-Loop Stochastic Algorithms for Difference of Max-Structured Weakly Convex Functions

In this paper, we study a class of non-smooth non-convex problems in the form of $\min_{x}[\max_{y\in Y}\phi(x, y) - \max_{z\in Z}\psi(x, z)]$, where both $\Phi(x) = \max_{y\in Y}\phi(x, y)$ and $\Psi(x)=\max_{z\in Z}\psi(x, z)$ are weakly convex functions, and $\phi(x, y), \psi(x, z)$ are strongly concave functions in terms of $y$ and $z$, respectively. It covers two families of problems that have been studied but are missing single-loop stochastic algorithms, i.e., difference of weakly convex functions and weakly convex strongly-concave min-max problems. We propose a stochastic Moreau envelope approximate gradient method dubbed SMAG, the first single-loop algorithm for solving these problems, and provide a state-of-the-art non-asymptotic convergence rate. The key idea of the design is to compute an approximate gradient of the Moreau envelopes of $\Phi, \Psi$ using only one step of stochastic gradient update of the primal and dual variables. Empirically, we conduct experiments on positive-unlabeled (PU) learning and partial area under ROC curve (pAUC) optimization with an adversarial fairness regularizer to validate the effectiveness of our proposed algorithms.

Paper

Similar papers

Peer review

Reviewer gunK6/10 · confidence 3/52024-06-30

Summary

This paper introduces a novel optimization problem termed Difference of Max-Structured Weakly Convex Functions (DMax). The DMax problem extends traditional frameworks such as difference-of-weakly-convex (DWC) optimization and weakly-convex-strongly-concave (WCSC) min-max optimization, which are widely utilized in the machine learning community. Using a Moreau envelope smoothing technique, the authors propose a stochastic algorithm called SMAG for optimizing DMax in non-smooth settings. This approach not only addresses the DMax problem but also applies to DWC and non-smooth WCSC min-max optimization scenarios, demonstrating comparable convergence rates. Experimental results validate the efficacy of their algorithms across various applications, including Positive-Unlabeled (PU) Learning and partial Area Under the Curve (AUC) optimization with an adversarial fairness regularizer.

Strengths

**Originality:** The paper introduces a new optimization problem termed Difference of Max-Structured Weakly Convex Functions (DMax), which extends existing frameworks like difference-of-weakly-convex (DWC) optimization and weakly-convex-strongly-concave (WCSC) min-max optimization. This extension demonstrates a significant level of originality in problem formulation within the field of optimization. By addressing a broader class of problems under the DMax framework, the paper contributes novel insights and methodologies that enrich the theoretical foundations of optimization in machine learning and related fields. **Quality:** The paper maintains high-quality standards across various aspects. It rigorously establishes the theoretical foundations of the DMax problem and proposes a Moreau envelope smoothing technique for optimization in non-smooth settings. The SMAG algorithm introduced is methodically developed and supported by well-defined lemmas and references, ensuring robustness and reliability in the proposed approach, though I don't check the proofs in detail. Furthermore, the experimental validation on applications such as Positive-Unlabeled Learning and partial AUC optimization with adversarial fairness regularization underscores the practical relevance and effectiveness of the proposed methods. **Clarity:** The clarity of the paper is a notable strength. It effectively communicates complex concepts and methodologies in a clear and accessible manner, making it understandable even for readers with limited expertise in the specific problem domain. The intuitive presentation of algorithms and key steps, coupled with clear explanations of referenced lemmas, enhances readability without sacrificing depth. This clarity not only facilitates comprehension but also promotes transparency in the theoretical derivations and experimental procedures, thereby bolstering the paper's overall impact. **Significance:** By addressing challenging optimization problems and providing practical algorithms validated in real-world scenarios, the paper significantly advances both theoretical understanding and practical applications in machine learning optimization.

Weaknesses

**Problem Statement and Novelty:** The paper introduces a new problem that extends traditional frameworks such as difference-of-weakly-convex (DWC) optimization and weakly-convex-strongly-concave (WCSC) min-max optimization, presenting the SMAG algorithm as a solution. However, the paper could benefit from clearer justification of how this generalization represents a significant departure from incremental advancements. This clarity is pivotal for establishing the theoretical contribution within a novel setting. **Complexity Analysis:** Upon examining the complexity results, it is apparent that in non-convex, non-smooth min-max problems, SMAG demonstrates complexity similar to Epoch-GDA under comparable assumptions (Table 2). This finding raises significant concerns about the broader implications for WCSC min-max optimization, particularly considering the emphasis on complexity as a contribution (Line 59). A clearer explanation is needed on how SMAG surpasses existing complexities in this domain, beyond the practical advantage of a single-loop method for tuning, to strengthen the paper's impact solely from a complexity perspective. **Algorithmic Technique:** The proof and algorithmic technique employed in this study resemble SBCD, utilizing the Moreau envelope to transform non-smooth problems into smooth ones, thereby enhancing accessibility. Hence, it is crucial to clearly delineate why SMAG achieves superior complexity results compared to SBCD or to identify any overlooked aspects that may explain this discrepancy. I welcome further discussion on these points to ensure a comprehensive evaluation. Thanks!

Questions

In the experiments section, the authors employ a linear classification model with hinge loss (Line 271). However, this setting appears to be a smooth one, potentially limiting the demonstration of the benefits of the authors' method in non-smooth DWC problems, despite SMAG showing superiority over other methods. Regarding the large variance in training curves observed, such as in FER 2013 (Figure 1), it raises questions about the stability and robustness of the method in practical applications. To enhance confidence in the results, I suggest exploring parameter tuning to achieve more stable and convincing outcomes.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper is theoretical, and I believe there is no need to confirm societal impact.

Reviewer QoFj5/10 · confidence 1/52024-07-14

Summary

This paper proposes a stochastic Moreau envelope approximate gradient method dubbed SMAG, the first single-loop algorithm for solving these problems, and provides a state-of-the-art non-asymptotic convergence rate. This paper achieves the best complexity of order $O(\epsilon^{-4})$. Futhermore, the agorithm of this paper only use a single loop which is easy to implement and tune the hyperparameters. A typo: Line 3, it should be $\Psi(x) = \max_z \psi(x, z)$.

Strengths

No

Weaknesses

No

Questions

This paper considers the problem of a special structure that is the difference of convex functions. However, in this paper, this special structure seems not be used. As I know, for the problem of difference of convex functions, there are several algorithms can achieve the faster convergenc rate of general non-convex functions. So, I wonder whether the results of this paper is useful enough.

Rating

5

Confidence

1

Soundness

3

Presentation

3

Contribution

3

Limitations

No

Reviewer 2iRa6/10 · confidence 3/52024-07-20

Summary

The paper considers the problem of minimizing a difference of maximum of two functions, under various regularity assumptions. Two important settings are difference of weakly convex functions and weakly convex strongly concave min-max problems. The authors propose a single loop algorithm with a convergence rate (defined in an appropriate sense) of $\epsilon^{-4}$. The authors apply their method to positive-unlabeled (PU) learning and partial area under ROC curve (pAUC) optimization with an adversarial fairness regularizer, and compare against existing methods.

Strengths

1. The paper is generally well written. The authors motivate the technical challenges well and how to get around them rather. 2. Getting single loop algorithms with the Moreau smoothing technique is challenging. The authors make important contributions in this regard. 3. The experimental results are encouraging. The authors show consistent improvement in their applications over the baselines.

Weaknesses

1. (Writing) The authors should distinguish if the difference/improvement from previous work are "double loop vs single loop", and/or faster rates and/or weaker assumptions. For the DWC setting, it seems it is "double loop vs single loop", and/or faster rates and weaker assumptions, where as for WCSC, it is "double loop vs single loop". 2. (Writing) The authors should mention the notion of optimalities in various settings ($\epsilon$ nearly critical and $\epsilon$ near stationarity) eraly on, and remark that they are standard.

Questions

1. Are the proof ideas useful to give single loop algorithms for non-smooth convex minimization via Moreau smoothing, with optimal rates? 2. Optimality: Is the $\epsilon^{-4}$ rate achieved optimal in these settings?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer jGsv5/10 · confidence 2/52024-07-30

Summary

The paper studies a class of optimization problem named DMax, i.e., to minimize a loss that is a difference of two max functions. This problem covers both difference of weakly-convex optimization and weakly-convex strongly-concave minimax optimization. Existing algorithms require double-loop structure to solve the inner problem within certain accuracy. This paper proposes a simple single-loop stochastic algorithm that achieves state-of-the-art convergence rates. Empirical experiments demonstrate the effectiveness of the proposed algorithm.

Strengths

The paper considers a class of difference of max-structured weakly-convex optimization problems, which can be interesting and novel in its problem fomulation. The proposed algorithm is simple to implement as it is single-loop structured and also clear to understand. The convergence rate matches exsiting works that use more complicated structures of the algorithms.

Weaknesses

1. There are some typos that may affect readibility. (1) Is the definition of $\Psi$ in line 3 of the abstract correct? Shouldn't it be $\max_z\psi$? (2) I guess that it should be "Single-Loop" in the title? (3) In line 27, there is an empty citation for adversarial learning. (4) In line 49, "sufficient to to". (5) "covnex" in line 4 of Table 2. (6) Should it be $\partial_x \psi$ in line 4 of Algorithm 1 and line 3 of Algorithm 2? 2. Although the problem fomulation is interesting in optimization theory, the current applications seem to be limited to only applications of the special cases DWC and WCSC min-max optimization. Therefore, the problem can be a little bit artificial in the sense that these applications do not necessarily require such a formulation. It can be done by studying DWC and min-max separately, and the major benefit is that DMax provides a unified way. Are there examples that DMax gives more applications beyond DWC and min-max optimization? 3. (Minor) Although I like the proposed algorithm, which is simple and clear to understand, it seems like all the technical steps exist before. The paper combines these known results and uses it for a new class of problem. For example, it is well-known that single-loop updates can give the same rate for strongly-concave problems compared to double-loop algorithms that first solve the inner-problems to a required accuracy. It is also known that weakly-convex problems can be handled using Moreau envelope and usually enjoy the same rate as smooth problems. As a result, the convergence rates and theoretical analysis in the paper are not surprising. I am the emergence reviewer. I do not carefully check all the technical details and proofs in the paper. However, the algorithmic ideas and analysis outline look good to me. I will be on the positive side with a low confidence rate.

Questions

See weaknesses.

Rating

5

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

See weaknesses.

Reviewer gunK2024-08-10

Thank you for the clear response. I have slightly increased my score, particularly for the clarification of the main difference between the analysis of SMAG and SBCD, which serves as the new proof technique.

Reviewer jGsv2024-08-12

Thanks for your response! I have no additional problems. I will keep my score, which is on the positive side. I am willing to support the acceptance of the paper.

Authorsrebuttal2024-08-12

Thank you for acknowledging our responses and the positive feedback.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC