Specialization, Adaptation and Optimization in Dilute Random Attractor Neural Networks

Abstract

In a biological context one is used to examples in which the properties of organisms are adapted to provide efficient operation in their particular environments, with otherwise similar systems differing in appropriate details when their environments are different. In this talk we demonstrate a similar phenomenon in some simple neural networks. In particular, we consider the performance of dilute attractor neural networks for associative memory with respect to optimal performances as characterized by two different measures, retrieval overlap and size of the basin from which retrieval is possible. We believe however that our conclusions are more broadly applicable¹.