Robust oscillations and bifurcations in cellular neural networks

Abstract

A software package for studying discrete time neural networks dynamics is presented as an efficient analysis tool having also capabilities for designing special purpose cellular neural networks (CNN's) such as period controlled oscillators, noise generators and chaotic based systems. Based on an information theory approach, an entropy associated with a given structure of the network was defined as a global descriptor of the system dynamics. Any kind of discrete time neural model with any size is allowed, including CNN's as a particular case. By performing two or one dimensional analysis trough the weights space, some new properties of the opposite-template CNN's were discovered, e.g. the existence of robustness domains which means that small controllable change in weights values implies no change in the network's entropy. Using this software package, robust chaotic networks capable of generating white-noise like signals were also discovered.<>