Convergence of Adam for Non-convex Objectives: Relaxed Hyperparameters and Non-ergodic Case

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: He, Meixuan, Liang, Yuqing, Liu, Jinlan, Xu, Dongpo
Format: Preprint
Veröffentlicht: 2023
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866929707926159360
author He, Meixuan
Liang, Yuqing
Liu, Jinlan
Xu, Dongpo
author_facet He, Meixuan
Liang, Yuqing
Liu, Jinlan
Xu, Dongpo
contents Adam is a commonly used stochastic optimization algorithm in machine learning. However, its convergence is still not fully understood, especially in the non-convex setting. This paper focuses on exploring hyperparameter settings for the convergence of vanilla Adam and tackling the challenges of non-ergodic convergence related to practical application. The primary contributions are summarized as follows: firstly, we introduce precise definitions of ergodic and non-ergodic convergence, which cover nearly all forms of convergence for stochastic optimization algorithms. Meanwhile, we emphasize the superiority of non-ergodic convergence over ergodic convergence. Secondly, we establish a weaker sufficient condition for the ergodic convergence guarantee of Adam, allowing a more relaxed choice of hyperparameters. On this basis, we achieve the almost sure ergodic convergence rate of Adam, which is arbitrarily close to $o(1/\sqrt{K})$. More importantly, we prove, for the first time, that the last iterate of Adam converges to a stationary point for non-convex objectives. Finally, we obtain the non-ergodic convergence rate of $O(1/K)$ for function values under the Polyak-Lojasiewicz (PL) condition. These findings build a solid theoretical foundation for Adam to solve non-convex stochastic optimization problems.
format Preprint
id arxiv_https___arxiv_org_abs_2307_11782
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Convergence of Adam for Non-convex Objectives: Relaxed Hyperparameters and Non-ergodic Case
He, Meixuan
Liang, Yuqing
Liu, Jinlan
Xu, Dongpo
Optimization and Control
Machine Learning
Numerical Analysis
Adam is a commonly used stochastic optimization algorithm in machine learning. However, its convergence is still not fully understood, especially in the non-convex setting. This paper focuses on exploring hyperparameter settings for the convergence of vanilla Adam and tackling the challenges of non-ergodic convergence related to practical application. The primary contributions are summarized as follows: firstly, we introduce precise definitions of ergodic and non-ergodic convergence, which cover nearly all forms of convergence for stochastic optimization algorithms. Meanwhile, we emphasize the superiority of non-ergodic convergence over ergodic convergence. Secondly, we establish a weaker sufficient condition for the ergodic convergence guarantee of Adam, allowing a more relaxed choice of hyperparameters. On this basis, we achieve the almost sure ergodic convergence rate of Adam, which is arbitrarily close to $o(1/\sqrt{K})$. More importantly, we prove, for the first time, that the last iterate of Adam converges to a stationary point for non-convex objectives. Finally, we obtain the non-ergodic convergence rate of $O(1/K)$ for function values under the Polyak-Lojasiewicz (PL) condition. These findings build a solid theoretical foundation for Adam to solve non-convex stochastic optimization problems.
title Convergence of Adam for Non-convex Objectives: Relaxed Hyperparameters and Non-ergodic Case
topic Optimization and Control
Machine Learning
Numerical Analysis
url https://arxiv.org/abs/2307.11782