Single-loop Projection-free and Projected Gradient-based Algorithms for Nonconvex-concave Saddle Point Problems with Bilevel Structure

TL;DR

This paper introduces novel single-loop projection-free and projected gradient algorithms for nonconvex-concave saddle point problems with bilevel structure, applicable to machine learning tasks like robust multi-task learning, with improved iteration complexities.

Contribution

It proposes the first single-loop, projection-free and projected gradient algorithms for bilevel saddle point problems, extending applicability beyond strongly concave cases.

Findings

The projection-free algorithm requires $ ilde{O}(rac{1}{ ext{epsilon}^4})$ iterations for an $ ext{epsilon}$-stationary point.

The gradient-based algorithm achieves $ ilde{O}(rac{1}{ ext{epsilon}^5})$ iterations, improved to $ ilde{O}(rac{1}{ ext{epsilon}^4})$ under strong concavity.

Experiments demonstrate the superiority of the proposed methods over state-of-the-art algorithms.

Abstract

In this paper, we explore a broad class of constrained saddle point problems with a bilevel structure, wherein the upper-level objective function is nonconvex-concave and smooth over compact and convex constraint sets, subject to a strongly convex lower-level objective function. This class of problems finds wide applicability in machine learning, encompassing robust multi-task learning, adversarial learning, and robust meta-learning. Our study extends the current literature in two main directions: (i) We consider a more general setting where the upper-level function is not necessarily strongly concave or linear in the maximization variable. (ii) While existing methods for solving saddle point problems with a bilevel structure are projection-based algorithms, we propose a one-sided projection-free method employing a linear minimization oracle. Specifically, by utilizing regularization and nested approximation techniques, we introduce a novel single-loop one-sided projection-free algorithm, requiring $\cO(\epsilon^{-4})$ iterations to attain an $\epsilon$-stationary solution, moreover, when the objective function in the upper-level is linear in the maximization component, our result improve to $\cO(\epsilon^{-3})$. Subsequently, we develop an efficient single-loop fully projected gradient-based algorithm capable of achieving an $\epsilon$-stationary solution within $\cO(\epsilon^{-5})$ iterations. This result improves to $\cO(\epsilon^{-4})$ when the upper-level objective function is strongly concave in the maximization component. Finally, we tested our proposed methods against the state-of-the-art algorithms for solving a robust multi-task regression problem to showcase the superiority of our algorithms.