Title: Convergence in Min-Max Optimization

Kevin A. Lai

Algorithms, Combinatorics, and Optimization

School of Computer Science

Georgia Institute of Technology

http://kevinalai.com/

Date:  Thursday, March 12th, 2020

Time:  10:30 am - 11:30 pm (ET)

Location: Klaus 2100

Committee:

Dr. Jacob Abernethy (Advisor, School of Computer Science, Georgia Institute of Technology)

Dr. Rachel Cummings (School of Industrial and Systems Engineering, Georgia Institute of Technology)

Dr. Jamie Morgenstern (School of Computer Science & Engineering, University of Washington)

Dr. Sebastian Pokutta (Institute of Mathematics, Technische Universität Berlin)

Dr. Mohit Singh (School of Industrial & Systems Engineering, Georgia Institute of Technology)

Read​er: Dr. Rachel Cummings

Abstract:

Many exciting recent developments in machine learning, such as Generative Adversarial Networks (GANs) and Deepmind's AlphaStar, rely on finding an equilibrium of a game as part of training. This process can be described as a min-max optimization problem. While min-max optimization has a long history of study, the difficulty of training systems such as GANs has revealed important gaps in our understanding of convergence in games. In this talk, we will cover several recent results that help fill some of these gaps. We cover results on fast convergence of fictitious play, last-iterate convergence, and new higher-order methods for min-max problems. This talk is based on joint work with Jacob Abernethy, Brian Bullins, and Andre Wibisono.