High-capacity exponential associative memories

Abstract

A generalized associative memory model with potentially high capacity is presented. A memory of this kind with M stored vectors of length N, can be implemented with M nonlinear neurons, N ordinary thresholding neurons, and 2MN binary synapses. It is shown that special cases of this model include the Hopfield and high-order correlation memories. A special case of the model, based on a neuron which can implement the subthreshold region, is presented. The authors analyze the capacity of this exponentially associative memory and show that it scales exponentially with N. In any practical realization, however, the dynamic range of the exponentiators is constrained. They show that the capacity for networks with fixed dynamic range exponential circuits is proportional to the dynamic range.