Analog implementations of artificial neural networks

Abstract

Analog implementations of artificial neural networks (ANNs) are discussed. Analog ANNs should be faster and smaller than digital implementations; however, the problem of the smaller dynamic range of analog storage must be addressed. The main problems are the required level of connectivity and long-term storage. With regard to connectivity, analog ANNs may be restricted to applications where only local connectivity is required, or where the number of neurons is small enough that essentially full connectivity can be achieved with VLSI. The problems with long-term storage are the complexity required and the resolution required for the backward error propagation (BEP) learning algorithm. The prospects of several different techniques for implementing analog ANNs are presented along with a brief survey of recent research results.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>