Skip to yearly menu bar Skip to main content

Workshop: Nonconvex Optimization for Machine Learning: Theory and Practice

Taming non-convexity via geometry

Suvrit Sra


In this talk, I will highlight some aspects of geometry and its role in optimization. In particular, I will talk about optimization problems whose parameters are constrained to lie on a manifold or in a specific metric space. These geometric constraints often make the problems numerically challenging, but they can also unravel properties that ensure tractable attainment of global optimality for certain otherwise non-convex problems.

We'll make our foray into geometric optimization via geodesic convexity, a concept that generalizes the usual notion of convexity to nonlinear metric spaces such as Riemannian manifolds. I will outline some of our results that contribute to g-convex analysis as well as to the theory of first-order g-convex optimization. I will mention several very interesting optimization problems where g-convexity proves remarkably useful. In closing, I will mention extensions to large-scale non-convex geometric optimization as well as key open problems.

Live content is unavailable. Log in and register to view live content