Bistable Gradient Neural Networks: Their Computational Properties

Abstract

A novel type of gradient attractor neural network is described that is characterized by bistable dynamics of nodes and their linear coupling. In contrast to the traditional perceptron-based neural networks which are plagued by spurious states it is found that this Bistable Gradient Network (BGN) is virtually free from spurious states. The consequences of this — greatly enhanced memory capacity, high speed of training and perfect recall — are illustrated by and compared with a small Hopfield network.