Almost sure convergence rates of stochastic gradient methods under gradient domination

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Weissmann, Simon, Klein, Sara, Azizian, Waïss, Döring, Leif
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866929760492322816
author Weissmann, Simon
Klein, Sara
Azizian, Waïss
Döring, Leif
author_facet Weissmann, Simon
Klein, Sara
Azizian, Waïss
Döring, Leif
contents Stochastic gradient methods are among the most important algorithms in training machine learning problems. While classical assumptions such as strong convexity allow a simple analysis they are rarely satisfied in applications. In recent years, global and local gradient domination properties have shown to be a more realistic replacement of strong convexity. They were proved to hold in diverse settings such as (simple) policy gradient methods in reinforcement learning and training of deep neural networks with analytic activation functions. We prove almost sure convergence rates $f(X_n)-f^*\in o\big( n^{-\frac{1}{4β-1}+ε}\big)$ of the last iterate for stochastic gradient descent (with and without momentum) under global and local $β$-gradient domination assumptions. The almost sure rates get arbitrarily close to recent rates in expectation. Finally, we demonstrate how to apply our results to the training task in both supervised and reinforcement learning.
format Preprint
id arxiv_https___arxiv_org_abs_2405_13592
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Almost sure convergence rates of stochastic gradient methods under gradient domination
Weissmann, Simon
Klein, Sara
Azizian, Waïss
Döring, Leif
Machine Learning
Optimization and Control
Stochastic gradient methods are among the most important algorithms in training machine learning problems. While classical assumptions such as strong convexity allow a simple analysis they are rarely satisfied in applications. In recent years, global and local gradient domination properties have shown to be a more realistic replacement of strong convexity. They were proved to hold in diverse settings such as (simple) policy gradient methods in reinforcement learning and training of deep neural networks with analytic activation functions. We prove almost sure convergence rates $f(X_n)-f^*\in o\big( n^{-\frac{1}{4β-1}+ε}\big)$ of the last iterate for stochastic gradient descent (with and without momentum) under global and local $β$-gradient domination assumptions. The almost sure rates get arbitrarily close to recent rates in expectation. Finally, we demonstrate how to apply our results to the training task in both supervised and reinforcement learning.
title Almost sure convergence rates of stochastic gradient methods under gradient domination
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2405.13592