Controlled Learning of Pointwise Nonlinearities in Neural-Network-Like Architectures

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Unser, Michael, Goujon, Alexis, Ducotterd, Stanislas
Format: Preprint
Veröffentlicht: 2024
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866909555592527872
author Unser, Michael
Goujon, Alexis
Ducotterd, Stanislas
author_facet Unser, Michael
Goujon, Alexis
Ducotterd, Stanislas
contents We present a general variational framework for the training of freeform nonlinearities in layered computational architectures subject to some slope constraints. The regularization that we add to the traditional training loss penalizes the second-order total variation of each trainable activation. The slope constraints allow us to impose properties such as 1-Lipschitz stability, firm non-expansiveness, and monotonicity/invertibility. These properties are crucial to ensure the proper functioning of certain classes of signal-processing algorithms (e.g., plug-and-play schemes, unrolled proximal gradient, invertible flows). We prove that the global optimum of the stated constrained-optimization problem is achieved with nonlinearities that are adaptive nonuniform linear splines. We then show how to solve the resulting function-optimization problem numerically by representing the nonlinearities in a suitable (nonuniform) B-spline basis. Finally, we illustrate the use of our framework with the data-driven design of (weakly) convex regularizers for the denoising of images and the resolution of inverse problems.
format Preprint
id arxiv_https___arxiv_org_abs_2408_13114
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Controlled Learning of Pointwise Nonlinearities in Neural-Network-Like Architectures
Unser, Michael
Goujon, Alexis
Ducotterd, Stanislas
Machine Learning
Functional Analysis
We present a general variational framework for the training of freeform nonlinearities in layered computational architectures subject to some slope constraints. The regularization that we add to the traditional training loss penalizes the second-order total variation of each trainable activation. The slope constraints allow us to impose properties such as 1-Lipschitz stability, firm non-expansiveness, and monotonicity/invertibility. These properties are crucial to ensure the proper functioning of certain classes of signal-processing algorithms (e.g., plug-and-play schemes, unrolled proximal gradient, invertible flows). We prove that the global optimum of the stated constrained-optimization problem is achieved with nonlinearities that are adaptive nonuniform linear splines. We then show how to solve the resulting function-optimization problem numerically by representing the nonlinearities in a suitable (nonuniform) B-spline basis. Finally, we illustrate the use of our framework with the data-driven design of (weakly) convex regularizers for the denoising of images and the resolution of inverse problems.
title Controlled Learning of Pointwise Nonlinearities in Neural-Network-Like Architectures
topic Machine Learning
Functional Analysis
url https://arxiv.org/abs/2408.13114