Saved in:
Bibliographic Details
Main Authors: Goel, Gautam, Soltanolkotabi, Mahdi, Bartlett, Peter
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.01514
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918365740662784
author Goel, Gautam
Soltanolkotabi, Mahdi
Bartlett, Peter
author_facet Goel, Gautam
Soltanolkotabi, Mahdi
Bartlett, Peter
contents We study the training dynamics of gradient descent in a softmax self-attention layer trained to perform linear regression and show that a simple first-order optimization algorithm can converge to the globally optimal self-attention parameters at a geometric rate. Our analysis proceeds in two steps. First, we show that in the infinite-data limit the regression problem solved by the self-attention layer is equivalent to a nonconvex matrix factorization problem. Second, we exploit this connection to design a novel "structure-aware" variant of gradient descent which efficiently optimizes the original finite-data regression objective. Our optimization algorithm features several innovations over standard gradient descent, including a preconditioner and regularizer which help avoid spurious stationary points, and a data-dependent spectral initialization of parameters which lie near the manifold of global minima with high probability.
format Preprint
id arxiv_https___arxiv_org_abs_2603_01514
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Training Dynamics of Softmax Self-Attention: Fast Global Convergence via Preconditioning
Goel, Gautam
Soltanolkotabi, Mahdi
Bartlett, Peter
Machine Learning
We study the training dynamics of gradient descent in a softmax self-attention layer trained to perform linear regression and show that a simple first-order optimization algorithm can converge to the globally optimal self-attention parameters at a geometric rate. Our analysis proceeds in two steps. First, we show that in the infinite-data limit the regression problem solved by the self-attention layer is equivalent to a nonconvex matrix factorization problem. Second, we exploit this connection to design a novel "structure-aware" variant of gradient descent which efficiently optimizes the original finite-data regression objective. Our optimization algorithm features several innovations over standard gradient descent, including a preconditioner and regularizer which help avoid spurious stationary points, and a data-dependent spectral initialization of parameters which lie near the manifold of global minima with high probability.
title Training Dynamics of Softmax Self-Attention: Fast Global Convergence via Preconditioning
topic Machine Learning
url https://arxiv.org/abs/2603.01514