Universal approximation and model compression for radial neural networks

TL;DR

This paper introduces radial neural networks with norm-dependent activation functions, proves their universal approximation capabilities, and develops a lossless model compression method leveraging their symmetry properties.

Contribution

It extends neural network theory by establishing universal approximation for radial networks and proposes a novel, symmetry-based model compression technique.

Findings

Radial neural networks can approximate any continuous function on bounded domains.

A symmetry-based lossless compression algorithm for radial networks is developed.

Optimization on compressed models is equivalent to projected gradient descent on full models.

Abstract

We introduce a class of fully-connected neural networks whose activation functions, rather than being pointwise, rescale feature vectors by a function depending only on their norm. We call such networks radial neural networks, extending previous work on rotation equivariant networks that considers rescaling activations in less generality. We prove universal approximation theorems for radial neural networks, including in the more difficult cases of bounded widths and unbounded domains. Our proof techniques are novel, distinct from those in the pointwise case. Additionally, radial neural networks exhibit a rich group of orthogonal change-of-basis symmetries on the vector space of trainable parameters. Factoring out these symmetries leads to a practical lossless model compression algorithm. Optimization of the compressed model by gradient descent is equivalent to projected gradient descent for the full model.