Constructing Convergent Methods for Minimax Optimization Problems
ABSTRACT
Minimax optimization appears in a range of modern machine learning problems, including generative adversarial networks, neural optimal transport, and adversarial training. Unlike classical minimization, however, minimax optimization is still not as well understood theoretically, which helps explain why such problems can be difficult to solve reliably in practice. In this talk, I will give an overview of some of my recent work on minimax methods with provable convergence guarantees. Starting from the classical extragradient method, I will describe extensions that handle stochasticity and that can be used to find local minimax points in nonconvex-nonconcave problems.
