Tailoring the performance of attractor neural networks

Abstract

First, we study the effects of introducing training noise on the retrievalbehaviours of dilute attractor neural networks. We found that, ingeneral, training noise enhances associativity, but also reducesthe attractor overlap. At a narrow range of storage levels, however,the system exhibits re-entrant retrieval behaviour on increasingtraining noise. Secondly, we consider optimization of network performance,and subsequently the storage capacity, in the presence of retrievalnoise (temperature). This is achieved by adapting the network toan appropriate training overlap, which is determined self-consistentlyby the optimal attractor overlap. The maximum storage capacity deviatesfrom the storage capacity of the maximally stable network on increasingtemperature, and in the high temperature regime (T \ge 0.38 forGaussian noise, the Hebb-rule network yields the maximum storagecapacity. Our analysis demonstrates the principles of specializationand adaptation in neural networks.