View abstract

Session II.2 - Continuous Optimization

Thursday, June 15, 17:00 ~ 17:30

Leveraging "partial" smoothness for faster convergence in nonsmooth optimization

Davis Damek

Cornell University, USA   -   This email address is being protected from spambots. You need JavaScript enabled to view it.

Nonsmoothness and nonconvexity are significant challenges in large-scale optimization problems, such as training neural networks and solving inverse problems in computational imaging. Although first-order methods are commonly used to address these problems, they are often considered slow. In this presentation, we introduce a (locally) accelerated first-order method that violates this perception by solving "generic" nonsmooth equations at a superlinear rate. The central insight is that nonsmooth functions often exhibit "partial" smoothness, which can be leveraged through a novel generalization of Newton's method.

Joint work with Vasileios Charisopoulos (Cornell University, University of Chicago).

View abstract PDF