Learning and Forgetting in Generalized Brain-state-in-a-box (BSB) Neural Associative Memories

We propose learning and forgetting techniques for the generalized brain-state-in-a-box (BSB) based associative memories. A generalization of the BSB model allows each neuron to have its own bias and the synaptic weight matrix does not have to be symmetric. A pattern is learned by a memory if its noisy or an incomplete version presented to the memory is mapped back to this pattern. A pattern, previously stored, is forgotten or deleted from the memory if a stimulus that is a perturbed version of the pattern, when presented to the memory, is not mapped back to this pattern. In this paper we propose "on-line" memory storage and deletion methods using an iterative method of computing the pseudo-inverse of a given matrix. The proposed methods allow one to "add" or "delete" a memory pattern by updating, rather than recomputing from scratch, the current synaptic weight matrix in a single step. We first analyze the desired characteristics of neural network associative memories. After that, we review the existing methods for design of neural associative memories. Then we discuss the generalized BSB neural model and its possible function as an associative memory and proffer arguments in support of using such models for neural associative memories. In particular, the generalized BSB type models are easier to analyze, synthesize, and implement than other neural networks. The results obtained are illustrated by numerical examples. Copyright 1996 Elsevier Science Ltd