Smooth Games in Machine Learning Beyond GANs

This talk will discuss a wide spectrum of recent advances in machine learning using smooth games, beyond the phenomenal GANs. Such showcases include reinforcement learning, robust and adversarial machine learning, approximate Bayesian computation, maximum likelihood estimation in exponential family, and etc. We show that all of these classical machine learning tasks can be reduced to solving (non) convex-concave min-max optimization problems. Hence, it is of paramount importance to developing a good theoretical understanding and principled algorithms for min-max optimization. We will review some of the theory and algorithms for smooth games and variational inequalities in the convex regime and shed some light on their counterparts in the non-convex regime.