Saved in:
Bibliographic Details
Main Authors: Xie, Shuo, Mohamadi, Mohamad Amin, Li, Zhiyuan
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2410.08198
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912423252852736
author Xie, Shuo
Mohamadi, Mohamad Amin
Li, Zhiyuan
author_facet Xie, Shuo
Mohamadi, Mohamad Amin
Li, Zhiyuan
contents Adam outperforms SGD when training language models. Yet this advantage is not well-understood theoretically -- previous convergence analysis for Adam and SGD mainly focuses on the number of steps $T$ and is already minimax-optimal in non-convex cases, which are both $\widetilde{O}(T^{-1/4})$. In this work, we argue that the exploitation of nice $\ell_\infty$-geometry is the key advantage of Adam over SGD. More specifically, we give a new convergence analysis for Adam under novel assumptions that loss is smooth under $\ell_\infty$-geometry rather than the more common $\ell_2$-geometry, which yields a much better empirical smoothness constant for GPT-2 and ResNet models. Our experiments confirm that Adam performs much worse when the favorable $\ell_\infty$-geometry is changed while SGD provably remains unaffected. We also extend the convergence analysis to blockwise Adam under novel blockwise smoothness assumptions.
format Preprint
id arxiv_https___arxiv_org_abs_2410_08198
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Adam Exploits $\ell_\infty$-geometry of Loss Landscape via Coordinate-wise Adaptivity
Xie, Shuo
Mohamadi, Mohamad Amin
Li, Zhiyuan
Machine Learning
Adam outperforms SGD when training language models. Yet this advantage is not well-understood theoretically -- previous convergence analysis for Adam and SGD mainly focuses on the number of steps $T$ and is already minimax-optimal in non-convex cases, which are both $\widetilde{O}(T^{-1/4})$. In this work, we argue that the exploitation of nice $\ell_\infty$-geometry is the key advantage of Adam over SGD. More specifically, we give a new convergence analysis for Adam under novel assumptions that loss is smooth under $\ell_\infty$-geometry rather than the more common $\ell_2$-geometry, which yields a much better empirical smoothness constant for GPT-2 and ResNet models. Our experiments confirm that Adam performs much worse when the favorable $\ell_\infty$-geometry is changed while SGD provably remains unaffected. We also extend the convergence analysis to blockwise Adam under novel blockwise smoothness assumptions.
title Adam Exploits $\ell_\infty$-geometry of Loss Landscape via Coordinate-wise Adaptivity
topic Machine Learning
url https://arxiv.org/abs/2410.08198