High-Dimensional Analysis of Gradient Flow for Extensive-Width Quadratic Neural Networks

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Martin, Simon, Biroli, Giulio, Bach, Francis
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911377648517120
author Martin, Simon
Biroli, Giulio
Bach, Francis
author_facet Martin, Simon
Biroli, Giulio
Bach, Francis
contents We study the high-dimensional training dynamics of a shallow neural network with quadratic activation in a teacher-student setup. We focus on the extensive-width regime, where the teacher and student network widths scale proportionally with the input dimension, and the sample size grows quadratically. This scaling aims to describe overparameterized neural networks in which feature learning still plays a central role. In the high-dimensional limit, we derive a dynamical characterization of the gradient flow, in the spirit of dynamical mean-field theory (DMFT). Under l2-regularization, we analyze these equations at long times and characterize the performance and spectral properties of the resulting estimator. This result provides a quantitative understanding of the effect of overparameterization on learning and generalization, and reveals a double descent phenomenon in the presence of label noise, where generalization improves beyond interpolation. In the small regularization limit, we obtain an exact expression for the perfect recovery threshold as a function of the network widths, providing a precise characterization of how overparameterization influences recovery.
format Preprint
id arxiv_https___arxiv_org_abs_2601_10483
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle High-Dimensional Analysis of Gradient Flow for Extensive-Width Quadratic Neural Networks
Martin, Simon
Biroli, Giulio
Bach, Francis
Optimization and Control
Disordered Systems and Neural Networks
Machine Learning
We study the high-dimensional training dynamics of a shallow neural network with quadratic activation in a teacher-student setup. We focus on the extensive-width regime, where the teacher and student network widths scale proportionally with the input dimension, and the sample size grows quadratically. This scaling aims to describe overparameterized neural networks in which feature learning still plays a central role. In the high-dimensional limit, we derive a dynamical characterization of the gradient flow, in the spirit of dynamical mean-field theory (DMFT). Under l2-regularization, we analyze these equations at long times and characterize the performance and spectral properties of the resulting estimator. This result provides a quantitative understanding of the effect of overparameterization on learning and generalization, and reveals a double descent phenomenon in the presence of label noise, where generalization improves beyond interpolation. In the small regularization limit, we obtain an exact expression for the perfect recovery threshold as a function of the network widths, providing a precise characterization of how overparameterization influences recovery.
title High-Dimensional Analysis of Gradient Flow for Extensive-Width Quadratic Neural Networks
topic Optimization and Control
Disordered Systems and Neural Networks
Machine Learning
url https://arxiv.org/abs/2601.10483