SAPD+: An Accelerated Stochastic Method for Nonconvex-Concave Minimax ProblemsDownload PDF

Published: 31 Oct 2022, Last Modified: 16 Jan 2023NeurIPS 2022 AcceptReaders: Everyone
Keywords: saddle point problems, nonconvex optimization, stochastic gradient, accelerated methods
Abstract: We propose a new stochastic method SAPD+ for solving nonconvex-concave minimax problems of the form $\min\max\mathcal{L}(x,y)=f(x)+\Phi(x,y)-g(y)$, where $f,g$ are closed convex and $\Phi(x,y)$ is a smooth function that is weakly convex in $x$, (strongly) concave in $y$. For both strongly concave and merely concave settings, SAPD+ achieves the best known oracle complexities of $\mathcal{O}(L\kappa_y\epsilon^{-4})$ and $\mathcal{O}(L^3\epsilon^{-6})$, respectively, without assuming compactness of the problem domain, where $\kappa_y$ is the condition number, and $L$ is the Lipschitz constant. We also propose SAPD+ with variance reduction, which enjoys the best known oracle complexity of $\mathcal{O}(L\kappa_y^2\epsilon^{-3})$ for weakly convex-strongly concave setting. We demonstrate the efficiency of SAPD+ on a distributionally robust learning problem with a nonconvex regularizer and also on a multi-class classification problem in deep learning.
Supplementary Material: pdf
11 Replies