Harmonic analysis of neural networks

Abstract

Neural networks models have attracted a lot of interest in recent years mainly because there were perceived as a new idea for computing. These models can be described as a network in which every node computes a linear threshold function. One of the main difficulties in analyzing the properties of these networks is the fact that they consist of nonlinear elements. I will present a novel approach, based on harmonic analysis of Boolean functions, to analyze neural networks. In particular I will show how this technique can be applied to answer the following two fundamental questions (i) what is the computational power of a polynomial threshold element with respect to linear threshold elements? (ii) Is it possible to get exponentially many spurious memories when we use the outer-product method for programming the Hopfield model?

Files

Additional details