A Sparsely Connected Asymmetric Neural Network and Its Possible Application to the Processing of Transient Spatio-Temporal Signals

Abstract

We consider randomly connected neural networks with sparse asymmetric connections. Low connectivity is achieved by chosing a certain number of input and output connections for every cell. We show that rudimentary layered structures of cells automatically emerge. Then we examine the autonomous behaviour of this network and some modified models, both from a nonlinear dynamical and from a functional point of view. We not only study the occurring attractors of limit-cycle type by statistical means but also report on the observation of chaos-like attractors. We demonstrate the ability of the network to react to transient spatio-temporal patterns in real-time by showing that attractor transitions due to time correlations of the transient signals happen.