Saved in:
Bibliographic Details
Main Authors: Chen, Thomas, Ewald, Patrícia Muñoz
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2408.01517
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910005669658624
author Chen, Thomas
Ewald, Patrícia Muñoz
author_facet Chen, Thomas
Ewald, Patrícia Muñoz
contents We prove that the standard gradient flow in parameter space that underlies many training algorithms in deep learning can be continuously deformed into an adapted gradient flow which yields (constrained) Euclidean gradient flow in output space. Moreover, for the $L^{2}$ loss, if the Jacobian of the outputs with respect to the parameters is full rank (for fixed training data), then the time variable can be reparametrized so that the resulting flow is simply linear interpolation, and a global minimum can be achieved. For the cross-entropy loss, under the same rank condition and assuming the labels have positive components, we derive an explicit formula for the unique global minimum.
format Preprint
id arxiv_https___arxiv_org_abs_2408_01517
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Gradient flow in parameter space is equivalent to linear interpolation in output space
Chen, Thomas
Ewald, Patrícia Muñoz
Machine Learning
Artificial Intelligence
Mathematical Physics
Optimization and Control
62M45, 37C10
We prove that the standard gradient flow in parameter space that underlies many training algorithms in deep learning can be continuously deformed into an adapted gradient flow which yields (constrained) Euclidean gradient flow in output space. Moreover, for the $L^{2}$ loss, if the Jacobian of the outputs with respect to the parameters is full rank (for fixed training data), then the time variable can be reparametrized so that the resulting flow is simply linear interpolation, and a global minimum can be achieved. For the cross-entropy loss, under the same rank condition and assuming the labels have positive components, we derive an explicit formula for the unique global minimum.
title Gradient flow in parameter space is equivalent to linear interpolation in output space
topic Machine Learning
Artificial Intelligence
Mathematical Physics
Optimization and Control
62M45, 37C10
url https://arxiv.org/abs/2408.01517