Complex neural networks

Abstract

Abstract This paper proposes complex neural networks in the form of a perceptron and an autocorrelation associatron. Their properties are described and the results of computer simulation are presented. The discriminant power of the complex perceptron is nearly twice that of the real perceptron because linear separability associated with the (one‐layer) real perceptron corresponds to separation by quadratic surfaces for the complex perceptron. Two methods of increasing the memory capacity of the autocorrelation associatron by complexification are proposed, the first of which is equivalent to the method using a nonmonotonic output function proposed by Morita et al. [1]. Complex neural networks will be an effective and simple means to introduce phase differences among impulse trains and hence to study such temporal phenomena as synchronization of impulse trains.