Non-Smooth Weakly-Convex Finite-sum Coupled Compositional Optimization

This paper investigates new families of compositional optimization problems, called $\underline{\bf n}$on-$\underline{\bf s}$mooth $\underline{\bf w}$eakly-$\underline{\bf c}$onvex $\underline{\bf f}$inite-sum $\underline{\bf c}$oupled $\underline{\bf c}$ompositional $\underline{\bf o}$ptimization (NSWC FCCO). There has been a growing interest in FCCO due to its wide-ranging applications in machine learning and AI, as well as its ability to address the shortcomings of stochastic algorithms based on empirical risk minimization. However, current research on FCCO presumes that both the inner and outer functions are smooth, limiting their potential to tackle a more diverse set of problems. Our research expands on this area by examining non-smooth weakly-convex FCCO, where the outer function is weakly convex and non-decreasing, and the inner function is weakly-convex. We analyze a single-loop algorithm and establish its complexity for finding an $\epsilon$-stationary point of the Moreau envelop of the objective function. Additionally, we also extend the algorithm to solving novel non-smooth weakly-convex tri-level finite-sum coupled compositional optimization problems, which feature a nested arrangement of three functions. Lastly, we explore the applications of our algorithms in deep learning for two-way partial AUC maximization and multi-instance two-way partial AUC maximization, using empirical studies to showcase the effectiveness of the proposed algorithms.

Paper

Similar papers

Peer review

Reviewer FiRa6/10 · confidence 3/52023-07-05

Summary

This paper addresses the finite-sum coupled compositional optimization (FCCO) scenario, relaxing the requirement of Lipschitz gradient for the involved functions. Instead, they consider weakly convex functions with certain monotonicity conditions. The paper introduces new algorithms and provides oracle complexity guarantees for computing a point with an epsilon-gradient norm relative to the Moreau envelop. The authors also demonstrate the applicability of these methods to AUC maximization tasks.

Strengths

It is always beneficial and sometimes nontrivial to extend the algorithm implementation and analytic technique for gradient Lipschitz functions to that of more general weakly convex functions. The authors also demonstrate practical applicability of the proposed new algorithms.

Weaknesses

Since I'm not acquainted with the FCCO setting, I will concentrate on the broader technical aspects instead. * L137: I don't get the idea of the "monotonic property" of a multivariate mapping f here. What is the "input" referring here? Let the function be f: R^n -> R^m. Is the input here referring to a vector in R^n? If so, how can we define the "monotonicity" with respect to such a multivariate function? * The whole development in this paper is built on the weak convexity of F proved in Proposition 4.2 and 4.4. However, the arguments there are not transparent to me. Let me take the proof of Proposition 4.2 for example. (a) L809 first inequality: Why do we have v'(g(y) - g(x)) \geq v'( ... )? It seems that the authors assumed v <= 0 here as f is assumed non-increasing in L808 (though the term "non-increasing" is not clearly defined in my opinion). However, it is not clear to me why such a "monotonicity" of f will promise the nonnegativity of subgradients v. For a smooth function f, the claim (v <= 0) holds trivially. But for a general non-smooth weakly convex function, even with the "monotonic assumption", this is nontrivial and a formal proof is required. (b) L809 third inequality: should L_2C_1 be L_2C_1*sqrt(n), where n is the input dimension of f? Please check. If that is true, then the weak convexity parameter would be depend on the dimension n of the input to f. Minor: * L805: should v'(g(x) - g(y)) be v'(g(y) - g(x))? * L145: \epsilon > 0 should be 1 > \epsilon >= 0.

Questions

See weaknesses.

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

Yes.

Authorsrebuttal2023-08-17

Questions?

Dear Reviewer FiRa and AC, Given that the discussion period is ending soon, we would like to follow up if there are some other questions from the reviewer that need us to clarify. Thank you for providing valuable comments on our paper! Regards Authors

Reviewer FiRa2023-08-17

Thank the authors for the clarification, which addressed my concerns and filled in the missing details in the statement and proof. I'll increase my score to 6.

Reviewer GoBV6/10 · confidence 3/52023-07-06

Summary

The paper considers a class of finite-sum coupled compositional optimization (FCCO) problems and a class of tri-level finite-sum coupled compositional optimization (TCCO) problems. Under the setting of non-smooth weakly-convex FCCO, the paper establishes the complexity of a single-loop algorithm with the tool of Moreau envelop. The algorithm is then extended to solve the non-smooth weekly-convex TCCO. Numerical results on the two-way partial AUC maximization and multi-instance two-way partial AUC maximization are also reported.

Strengths

Convergence analysis of an single-loop stochastic algorithm for solving non-smooth weakly-convex FCCO is a nice contribution of the paper.

Weaknesses

1. There lacks motivation for considering non-smooth weakly-convex FCCO and TCCO. Some additional applications other than the two-way partial AUC maximization should be provided. 2. In the nonconvex setting, definitions need to be rigorously given/referred to. For example, $f$ at line 133 can be nonconvex, which definition of "subgradient" and "subdifferential" are the paper using? There some ways to define a "variance-reduced estimator", which definition is the paper using? 3. The experiments just partially support the theoretical results of the paper. Numerical results on the comparison of the evolution of the objective function over time when using the proposed algorithm and other known algorithms for solving one same non-smooth weakly-convex FCCO problem are expected. It would illustrate the importance of the obtained complexity.

Questions

Consider the TPAUC problem in Section 5.The corresponding NSWC FCCO problem of the regular setting and the corresponding NSWC TCCO problem for the MIL setting still have convex $f_i$. Could the complexity of the proposed algorithms be improved if the weakly-convex assumption of $f_i$ is replaced by convexity?

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

2 fair

Contribution

3 good

Limitations

n.n.

Reviewer evM37/10 · confidence 3/52023-07-12

Summary

This paper handles the problems of non-smooth weakly-convex compositional optimization. The first problem, referred to as FCCO (Finite-sum coupled compositional minimization), is given by $$ \min_{w \in \mathbb{R}^d} F(w) \triangleq \frac{1}{n} \sum_{i=1}^n f_i(\mathbb{E}_{\xi\sim \mathcal{D}_i}[g_i(w; \xi)])$$ The second problem, referred to as TCCO (Tri-level coupled compositional minimization), is given by, $$ \min_{w \in \mathbb{R}^d} F(w) \triangleq \frac{1}{n_1} \sum_{i=1}^{n_1} f_i\left(\frac{1}{n_2}\sum_{j=1}^{n_2} g_i( \underset{\xi \sim \mathcal{D}{i,j}}{\mathbb{E}} [h_{i,j}(w; \xi)])\right)$$ Here, the outer functions $f_i$ and $g_i$ are Lipschitz and weakly convex which has not been considered by previous works and significantly complicates the analysis. The authors propose two algorithms, SONX and SONT for the two problems respectively. For the first problem, FCCO, $u_{i,t}$ are the estimates of $g_i(w_t)$, $\forall i \in \{1,2,\ldots, n\}$, which are required for an unbiased estimate of the the complete function $F(w)$. In each step, a batch $B$ of coordinates from $\{1,2,\ldots, n\}$ is sampled and only the estimates in this batch, $u_{i,t}, i\in B$, are updated using a Variance-reduced estimator. Using the estimates, a single gradient step is computed for $w_t$. The estimator has been adapted from [1]. As the functions are weakly convex, the authors show convergence to an $\epsilon$-stationary point of the Moreau envelope of $F$ at a rate of $\mathcal{O}(\epsilon^{-6})$. The authors show that both problems can be mapped exactly to Two-way partial AUC (TPAUC) maximization of a parametrized network, even deep networks, with the TCCO problem mapping to a multi-instance learning version. TPAUC is the area under the ROC curve when False Positive rate $\leq \beta$ and True Positive Rate $\geq \alpha$. Over a set of medical datasets, the authors show that SONX and SONT out-perform baselines for TPAUC maximization as the baselines use poor approximations of TPAUC. **References** 1. Jiang et al 2022. Multi-block-single-probe variance reduced estimator for coupled compositional optimization

Strengths

- **Easy to Implement**: The algorithms are single-loop and use moderately sized batches which makes it easy to implement. Also, the algorithms can also be parallelized over the coordinates $i$. - **Better Rates**: In addition to being easy to implement, the algorithms achieve the best possible rates for this class of functions. Existing works obtain similar or worse rates under an easier problem setting or with multiple loops or using large batches. - **Mapping to TPAUC**: The authors describe the TPAUC problem comprehensively and show the exact mapping with the two optimization problems. - **Detailed presentation**: The convergence analysis, the mapping to TPAUC and the related works are very detailed which helps in understanding the paper.

Weaknesses

- **Experimental and Theoretical Baselines do not match**: For theory, the baselines for comparisons are algorithms for simpler problem settings, for instance with smoothness or convexity. For experiments, the baselines for TPAUC maximization are used which do not match the theoretical baselines. It is unclear, if some theoretical baseline can be applied to TPAUC and perform better. This should not be the case ideally but it has not been verified.

Questions

- Is there a lower bound for this problem setting, which can be used to verify if the rates are optimal and cannot be improved? - Why is the dependence on $\frac{n_1}{B_1}$ worse for SONT as compared to SONT?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

Described in Weaknesses

Reviewer mWSr6/10 · confidence 2/52023-07-24

Summary

The purpose of this paper is to introduce a new approach to solving a specific type of optimization problem called non-smooth weakly-convex finite-sum coupled compositional optimization (NSWC FCCO) problems. The authors specifically focus on a variation of FCCO problems where the outer function is weakly convex and non-decreasing, and the inner function is weakly convex. To address these problems, the authors propose two new algorithms. The first algorithm is designed for two-level NSWC FCCO problems and operates using a single-loop. The second algorithm is an extension of the first algorithm and is intended for tri-level NSWC FCCO problems. The authors provide a comprehensive analysis. They establish the complexity of the algorithms in terms of finding an $ε$-stationary point of the Moreau envelope of the objective function.

Strengths

The paper is of high quality as it provides a detailed and rigorous mathematical analysis of the proposed algorithms. The technique of using Moreau envelope is intuitive and interesting. I have not checked the proof line by line but it looks good. The authors also conduct extensive experiments to validate their theoretical findings. They compare the performance of their algorithms with other competitive methods across multiple datasets, demonstrating the effectiveness of their approach.

Weaknesses

The authors have presented new convergence analysis for FCCO and TCCO, which I acknowledge. However, it seems that other multistage algorithms have already achieved the optimal rate. Consequently, I would lower my evaluation score for the novelty aspect.

Questions

I acknowledge and value the authors' dedication in exploring the intricacies of the technical aspects. Nevertheless, I fail to find the application of maximizing two-way partial AUC particularly inspiring. Could you please demonstrate any other compelling applications?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

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

no

Reviewer SCt15/10 · confidence 5/52023-07-27

Summary

The manuscript studies a class of non-smooth non-convex compositional optimization problems in which the objective function is given in a form of finite-sum composition where the functions are assumed to be weakly convex. The authors present stochastic approximation algorithms to solve this class of problems and establish their sample complexity bounds.

Strengths

The authors provide the sample complexity bound of $O(1/\epsilon^6)$ for a single-loop algorithm applied to non-smooth weakly convex compositional problems.

Weaknesses

Given the existing works in the literature and assumptions made in the manuscript, the presented theoretical results are incremental.

Questions

1. In Table 1, the authors compare their results with the existing ones for smooth problems. However, dependence on $n$ is hidden and the reader cannot compare this dependency. Moreover, they need to add more existing results (mentioned on page 3) to the table with their assumptions so that there is a clear picture for comparison. 2. The choice of bath sizes appear in the complexity bounds, however, there is a very limited discussion on its role.

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

.

Reviewer SCt12023-08-17

Thanks for the clarification. I increased my rating to 5.

Reviewer evM32023-08-18

Response

Thanks for the detailed rebuttal, especially the clarification about the baseline SOTAs.

Reviewer mWSr2023-08-19

Thank you for the detailed reply, the paper looks more promising now and i will increase the score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC