Overview
Explore a 59-minute lecture on stochastic optimization algorithms with adaptive state-dependent variance for non-convex objective functions. Delve into the application of these methods in machine learning, where non-convex loss landscapes are prevalent. Learn about the dual-phase approach: an exploration phase followed by a faster converging algorithm in the basin of attraction. Discover how these techniques can be used independently, offering provable algebraic convergence rates optimal for the exploration phase. Examine numerical examples that complement the mathematical analysis, covering algorithms both with and without explicit use of gradient information. Gain insights from Björn Engquist's presentation at #ICBS2024, hosted by BIMSA, on achieving global convergence in stochastic optimization.
Syllabus
Björn Engquist: Global Convergence in Stochastic Optimization #ICBS2024
Taught by
BIMSA