Ashwin Pananjady (Georgia Tech)

Mar 2, 2022

Title and Abstract

Sharp convergence guarantees for iterative algorithms in random optimization problems

Iterative algorithms are the workhorses of modern statistical learning, and are widely used to fit large-scale, complex models to random data. While the choice of an algorithm and its hyperparameters determines both the speed and fidelity of the learning pipeline, it is common for this choice to be made heuristically, either by expensive trial-and-error or by comparing upper bounds on convergence rates of various candidate algorithms. Motivated by this, we develop a principled framework that produces sharp, iterate-by-iterate characterizations of solution quality for algorithms run with sample-splitting on a wide range of nonconvex model-fitting problems with Gaussian data. I will present the general framework and use it to derive concrete consequences for some popular algorithms in statistical settings, showcasing how sharp predictions can provide precise separations between the convergence behavior of algorithms while also revealing some nonstandard convergence phenomena. The talk will be based on collaborations with Kabir Chandrasekher, Mengqi Lou, and Christos Thrampoulidis.

Bio

Ashwin Pananjady is an Assistant Professor at Georgia Tech, with a joint appointment between the Schools of ISyE and ECE. His research currently focuses on the interplay of statistics and optimization in nonparametric ranking, nonconvex model-fitting, and reinforcement learning. He is a recipient of the Lawrence Brown Award from the Institute of Mathematical Statistics and the David Sakrison Memorial Prize from the EECS department at Berkeley, where his most significant accomplishment was to convince his (co-)advisor Martin Wainwright to walk up the hill to Cory Hall instead of taking the bus