Who: Swati Padmanabhan (Assistant Professor at the Department of Industrial and Systems Engineering of the University of Minnesota Twin Cities)
When: Thursday May 28th, 3:30-4:30pm
Where: Anderson 329
Speaking on: “A gradient sampling method for nonsmooth nonconvex optimization”
Many optimization methods rely on gradients, but gradients can be unreliable when the objective is nonsmooth: they may fail to exist, change abruptly, or point in misleading directions. This raises a basic question: what kind of algorithmic guarantees are still possible for minimizing general Lipschitz, nonsmooth, nonconvex functions? This talk will describe recent progress through the lens of Goldstein stationarity, a local certificate based on averaging gradients in a small neighborhood. The certificate leads to a geometric problem: finding a short vector in the convex hull of nearby gradients. Randomized gradient sampling makes this problem tractable for arbitrary Lipschitz functions, while a newer Frank-Wolfe-inspired viewpoint turns the same task into a parameter-free averaging procedure. The talk will conclude with an anytime algorithm that adaptively searches over neighborhood scales, balancing descent progress against increasingly local stationarity certificates. These results are based on joint works with Damek Davis, Dmitriy Drusvyatskiy, Yin Tat Lee, Guanghao Ye, Zhe Zhang, and Zitao Song.
As always, you don’t have to be a CS major to attend CS department events. Any members of the Carleton community with any interest at all in CS are more than welcome to show up!