QUALITATIVE LIMITATIONS INCURRED IN IMPLEMENTATIONS OF RECURRENT NEURAL NETWORKS

During the implementation process of artificial neural networks, devia tions from the desired ideal neural network are frequently introduced. These include parameter perturbations, transmission delays, and inter connection constraints. In the present article, we study the effects o f these realities of imperfection on the qualitative behavior of artif icial feedback neural networks. To accomplish this, we utilize a speci fic class of neural networks (Hopfield-like neural networks) with a sp ecific application (the realization of associative memories) as a vehi cle for our study. The principal issues which we address concern the e ffects of parameter perturbations, transmission delays, and interconne ction constraints on the accuracy and on the qualitative properties of the network memories.