Title: How to Escape Saddle Points Efficiently
|University of Waterloo
Abstract: We will discuss the 2017 paper named in the title of the talk, by Jin et al. This paper outlines a perturbed form of gradient descent that converges to second-order stationary points in a nearly “dimension-free” manor, with rates comparable to standard gradient descent. Specifically, if all saddle points for a given problem are non-degenerate, we will show that perturbed gradient descent can escape these saddle points almost for free. We will discuss applications where these saddle point assumptions are reasonable, and will conclude with a discussion on a novel characterization of saddle point geometry which made this result possible.