26
15

On minimal representations of shallow ReLU networks

Abstract

The realization function of a shallow ReLU network is a continuous and piecewise affine function f:RdRf:\mathbb R^d\to \mathbb R, where the domain Rd\mathbb R^{d} is partitioned by a set of nn hyperplanes into cells on which ff is affine. We show that the minimal representation for ff uses either nn, n+1n+1 or n+2n+2 neurons and we characterize each of the three cases. In the particular case, where the input layer is one-dimensional, minimal representations always use at most n+1n+1 neurons but in all higher dimensional settings there are functions for which n+2n+2 neurons are needed. Then we show that the set of minimal networks representing ff forms a CC^\infty-submanifold MM and we derive the dimension and the number of connected components of MM. Additionally, we give a criterion for the hyperplanes that guarantees that all continuous, piecewise affine functions are realization functions of appropriate ReLU networks.

View on arXiv
Comments on this paper