Vergangene Vorträge im Joint Analysis Seminar
In den vergangenen sechs Monaten haben keine Veranstaltungen stattgefunden.
Vergangene Vorträge im Oberseminar
Andrea Paudice (Aarhus University, online):
General Tail Bounds for Non-Smooth Stochastic Mirror Descent
We study the problem of minimizing a convex, non-smooth Lipschitz function over a convex domain when only noisy stochastic subgradient estimates are available. We analyze
the classical Stochastic Mirror Descent (SMD) algorithm and derive new tail bounds on its optimization error, for both the averaged and the last iterate. Our results extend existing
analyses - traditionally limited to light-tailed, sub-Gaussian noise - to heavier-tailed noise distributions. We specialize our general bounds to two important families of noise: one
with exponential tails and another with polynomial tails. Notably, our bounds for the averaged iterate reveal a distinct two-regime behavior, highlighting new insights into the
interplay between noise tails and convergence rates.
(Abstract ein-/ausblenden)
