On non-approximability of zero loss global L^2 minimizers by gradient descent in deep learning

DOI10.2298/TAM250121008CMaRDI QIDQ6945392zbMATH OpenFDO

Authors Thomas Chen, Patrícia Muñoz Ewald

Publication date 8 August 2025

Published in Theoretical and Applied Mechanics (Search for Journal in Brave)

Full work available at URL https://doi.org/10.2298/TAM250121008C

zbMATH Keywords

deep learninggeneric training dataunderparametrizationzero loss

Mathematics Subject Classification ID

Critical points and critical submanifolds in differential topology (57R70) Neural nets and related approaches to inference from stochastic processes (62M45)

Cites work

• Mathematical Aspects of Deep Learning
• Prevalence of neural collapse during the terminal phase of deep learning training

Cited in

(2)

• Gradient flow in parameter space is equivalent to linear interpolation in output space
• Interpretable global minima of deep ReLU neural networks on sequentially separable data

This page was built for publication: On non-approximability of zero loss global \(\mathcal{L}^2\) minimizers by gradient descent in deep learning

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6945392)