Harmonic Decompositions of Convolutional Networks

Part of Proceedings of the International Conference on Machine Learning 1 pre-proceedings (ICML 2020)

Bibtex »Metadata »Paper »

Authors

Meyer Scetbon, Zaid Harchaoui

Abstract

We present a description of function spaces and smoothness classes associated with convolutional networks from a reproducing kernel Hilbert space viewpoint. We establish harmonic decompositions of convolutional networks, that is expansions into sums of elementary functions of feature-representation maps implemented by convolutional networks. The elementary functions are related to the spherical harmonics, a fundamental class of special functions on spheres. These harmonic decompositions allow us to characterize the integral operators associated with convolutional networks, and obtain as a result risk bounds for convolutional networks which highlight their behavior in high dimensions.