PRUNING ALGORITHM IN A DMA MODEL OF NEURAL NETWORKS

Abstract

Conventional network pruning techniques are uniquely established for multilayer neural networks in order to improve generalization or to prune processing units. The research presented in this paper is dealing with the modification of the architecture of an auto-associative neural network performed by progressive adaptation of the external stimuli. The proposed network pruning algorithm was obtained after an exhaustive experimenting with the Distributed Memory and Amnesia DMA model. The results of the experiments performed over a network with growing and constant architecture are presented. Finally, the dynamic shrink of the network by discarding the neurons one by one or simultaneously is compared.