BICONN: A binary competitive neural network

Abstract

In this paper a competitive neural network with binary synaptic weights is proposed. The aim of this network is to cluster or categorize binary input data. The neural network uses a learning mechanism based on activity levels that generates new binary synaptic weights that evolve toward medianoids of the clusters or categorizes that are being formed by the process units of the network, since the medianoid is the better representation of a cluster for binary data when the Hamming distance is used. The proposed model has been applied to codebook generation in vector quantization (VQ) for binary fingerprint image compression. The binary neural network find a set of representative vectors (codebook) for a given training set minimizing the average distortion. © Springer-Verlag Berlin Heidelberg 2003.