Solving hidden monotone variational inequalities with surrogate losses

I will talk about our recent work on using surrogate losses to solve variational inequality (VI) problems. Surrogate losses consider optimization in function space (or model predictions) instead of directly over model parameters and have been recently used in understanding and designing new RL policy gradient methods. I will discuss how we can generalize this idea beyond scalar minimization and take advantage of hidden monotone structure to establish global convergence guarantees. Theoretical challenges using surrogate losses in VI problems will also be discussed, along with applications to min-max and RL policy evaluation.