A Dual Certificate Approach to Sparsity in Infinite-Width Shallow Neural Networks

TL;DR

This paper analyzes the sparsity of solutions in infinite-width shallow ReLU neural networks trained with TV regularization, providing theoretical guarantees and conditions for sparsity, uniqueness, and convergence based on duality and data geometry.

Contribution

It introduces a dual certificate framework to establish sparsity and uniqueness of solutions, linking the structure to data-induced hyperplane arrangements and analyzing convergence under low noise.

Findings

Support size of solutions is finite and bounded by data geometry.

Solutions remain sparse with a fixed number of Dirac deltas under low noise.

Convergence rate depends linearly on noise and regularization parameter.

Abstract

In this paper, we study total variation (TV)-regularized training of infinite-width shallow ReLU neural networks, formulated as a convex optimization problem over measures on the unit sphere. Our approach leverages the duality theory of TV-regularized optimization problems to establish rigorous guarantees on the sparsity of the solutions to the training problem. Our analysis further characterizes how and when this sparsity persists in a low noise regime and for small regularization parameter. The key observation that motivates our analysis is that, for ReLU activations, the associated dual certificate is piecewise linear in the weight space. Its linearity regions, which we name dual regions, are determined by the activation patterns of the data via the induced hyperplane arrangement. Taking advantage of this structure, we prove that, on each dual region, the dual certificate admits at most one extreme value. As a consequence, the support of any minimizer is finite, and its cardinality can be bounded from above by a constant depending only on the geometry of the data-induced hyperplane arrangement. Then, we further investigate sufficient conditions ensuring uniqueness of such sparse solution. Finally, under a suitable non-degeneracy condition on the dual certificate along the boundaries of the dual regions, we prove that in the presence of low label noise and for small regularization parameter, solutions to the training problem remain sparse with the same number of Dirac deltas. Additionally, their location and the amplitudes converge, and, in case the locations lie in the interior of a dual region, the convergence happens with a rate that depends linearly on the noise and the regularization parameter.