Аннотация:
The talk will focus on limit theorems for stochastic gradient descent
(SGD) and on how the choice of step size affects its convergence. We
will begin by discussing the convergence of deterministic gradient
descent with a constant step size, then review the classical
Robbins–Monro result for SGD, and finally turn to several results from
our joint work [1] concerning the convergence of SGD with a constant
step size that tends to zero.
D. Dudukalov, A. Logachov, V. Lotov, T. Prasolov, E. Prokopenko & A. Tarasenko, Convergence, Sticking and Escape: Stochastic Dynamics Near Critical Points in SGD, arXiv preprint, 2025, arXiv: 2505.18535