TILOS-Optimization for ML and AI Seminar: Implicit bias results for Muon, Adam, and Friends
Matus Telgarsky, New York University
Abstract: This talk will give both an empirical overview and a few simple bonds controlling the optimization path, or implicit bias, of modern optimization methods such as Adam and Muon (and Friends). The talk will begin with empirical results demonstrating the implicit bias phenomenon with shallow networks and also transformers combined with chain-of-thought. The talk will then briefly survey a few mathematical implicit bias analyses of nonlinear networks, which unfortunately do not carry through to transformers. As such, the talk concludes with a technical portion presenting another approach to analyzing these optimization methods in the linear case, providing generic implicit bias results for them, and empirically demonstrating hope that this particular methodology can carry over to the nonlinear case.
Matus Telgarsky is an Associate Professor of Computer Science at the Courant Institute of Math at NYU, specializing in deep learning theory. The highlight of his academic career was completing a PhD under Sanjoy Dasgupta at UC San Diego. Adventures since then include co-chairing the Midwest ML Symposium in 2017 with Po-Ling Loh, and chairing two semester-long Simons Institute Programs at UC Berkeley. Accolades include a 2018 NSF Career Award and delivering a COLT 2025 keynote.