Abstract

A new three-layer neural network is proposed. A special structure of connections between the layer of input nodes and the intermediate layer forms the memory matrix of the network. The intermediate layer is composed of neuronlike threshold elements and is totally interconnected by feedback loops. The weights of the lateral inhibition connections are determined by the cross correlations of the stored vectors. Programming the interconnections between the intermediate layer and an output slab allows any arbitrary output for any input. The network was designed to perform as a content-addressable memory or, for example, as a symbolic substitution system with a winner-take-all function, which selects the neuron with the largest activity. Such an architecture has been shown^1,2 to be a better neural network processor than the widely known and discussed Hopfield model.

© 1988 Optical Society of America