ROBUSTNESS ANALYSIS AND DESIGN OF A CLASS OF NEURAL NETWORKS WITH SPARSE INTERCONNECTING STRUCTURE

We first conduct an analysis of the robustness Properties of-a class o f neural networks with applications to associative memories. Specifica lly, for a network with nominal parameters which stores a set of desir ed bipolar memories, we establish sufficient conditions under which th e same set of bipolar memories is also stored in the network with pert urbed parameters. This result enables us to establish a synthesis proc edure for neural networks whose stored memories are invariant under pe rturbations. Our synthesis procedure is capable of generating artifici al neural networks with prespecified sparsity constraints (on the inte rconnecting structure) and with nonsymmetric and symmetric interconnec tion matrices. To demonstrate the applicability of the present results , we consider several specific examples.