DYNAMICS OF HIGHLY DILUTED NETWORKS OF CONTINUOUS NEURONS

Abstract

We study the dynamics of the maximally stable network of analog neurons with a parallel update rule. For this type of network we determine the size of the basins of attraction and the effect of random destruction of synapses on the size of the basins of attraction and the attractor overlap. We present analytical and numerical results showing the significant increase of the wide retrieval regime in comparison with binary networks. On the other hand, we find that, in contradistinction to binary networks, our analog network fails to perform as an autoassociative memory if the percentage of cancelled synapses is larger than a certain critical value.