Feature learning is decoupled from generalization in high capacity neural networks

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Göring, Niclas Alexander, London, Charles, Erturk, Abdurrahman Hadi, Mingard, Chris, Nam, Yoonsoo, Louis, Ard A.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909706390339584
author Göring, Niclas Alexander
London, Charles
Erturk, Abdurrahman Hadi
Mingard, Chris
Nam, Yoonsoo
Louis, Ard A.
author_facet Göring, Niclas Alexander
London, Charles
Erturk, Abdurrahman Hadi
Mingard, Chris
Nam, Yoonsoo
Louis, Ard A.
contents Neural networks outperform kernel methods, sometimes by orders of magnitude, e.g. on staircase functions. This advantage stems from the ability of neural networks to learn features, adapting their hidden representations to better capture the data. We introduce a concept we call feature quality to measure this performance improvement. We examine existing theories of feature learning and demonstrate empirically that they primarily assess the strength of feature learning, rather than the quality of the learned features themselves. Consequently, current theories of feature learning do not provide a sufficient foundation for developing theories of neural network generalization.
format Preprint
id arxiv_https___arxiv_org_abs_2507_19680
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Feature learning is decoupled from generalization in high capacity neural networks
Göring, Niclas Alexander
London, Charles
Erturk, Abdurrahman Hadi
Mingard, Chris
Nam, Yoonsoo
Louis, Ard A.
Machine Learning
Neural networks outperform kernel methods, sometimes by orders of magnitude, e.g. on staircase functions. This advantage stems from the ability of neural networks to learn features, adapting their hidden representations to better capture the data. We introduce a concept we call feature quality to measure this performance improvement. We examine existing theories of feature learning and demonstrate empirically that they primarily assess the strength of feature learning, rather than the quality of the learned features themselves. Consequently, current theories of feature learning do not provide a sufficient foundation for developing theories of neural network generalization.
title Feature learning is decoupled from generalization in high capacity neural networks
topic Machine Learning
url https://arxiv.org/abs/2507.19680