Symmetries and learning in neural network models

Abstract

I consider learning in neural network models and demonstrate how global properties can be derived from the characteristics of the local synaptic modification rules. I examine, in detail, the case of the Hopfield model of associative memory with Hebbian learning and show how the configuration space is partitioned into orbits of points of equal behavior, yielding a description of the structure of all the stable points. Calculations for a small-sized example are given.