Gradient flow in parameter space is equivalent to linear interpolation in output space

DOI10.1016/J.GEOMPHYS.2026.105765MaRDI QIDQ6858039zbMATH OpenFDO

Authors Thomas Chen, Patrícia Muñoz Ewald

Publication date 13 February 2026

Published in Journal of Geometry and Physics (Search for Journal in Brave)

Full work available at URL https://doi.org/10.1016/J.GEOMPHYS.2026.105765

zbMATH Keywords

constrained dynamicsdeep learninggradient flowshomotopy equivalencesneural tangent kernel

Mathematics Subject Classification ID

Flows related to symplectic and contact structures (53E50)

Cites work

• Functional dimension of feedforward ReLU neural networks
• Generalized Inverses of Linear Transformations
• Global minima of overparameterized neural networks
• scientific article; zbMATH DE number 7306852 (Why is no real title available?)
• Information geometry
• Information geometry and its applications
• Natural gradient via optimal transport
• On non-approximability of zero loss global \(\mathcal{L}^2\) minimizers by gradient descent in deep learning
• Ordinary differential equations and dynamical systems
• Prevalence of neural collapse during the terminal phase of deep learning training

Cited in

(1)

• Geometric structure of shallow neural networks and constructive \(\mathcal{L}^2\) cost minimization

This page was built for publication: Gradient flow in parameter space is equivalent to linear interpolation in output space

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6858039)