Neural network model using a normalized inner product as a measure of similarity

A three-layer neural network model is proposed. The input and output layers of neurons consist of nodes to interconnect the inputs and outputs to the intermediate layer. The intermediate layer of neurons is a modified Grossberg slab with feedback whose function is a winner-take-all operation. The network improves on a previously proposed model by taking into account the number of nonzero elements in the binary stored vectors. The network can be used as a content-addressable memory or as a symbolic substitution system which yields an arbitrarily defined output for any input. Computer simulations using the network as an autoassociative content-addressable memory show improved performance over the previous model.<>