Competitive model to classify unknown data into hierarchical clusters through unsupervised learning
Y. Meki, Toshiki Kindo, Hiroaki Kurokawa, Iwao Sasase · 2002
We present a multi layer competitive model to classify unknown data into hierarchical clusters through unsupervised learning. The first layer of the model, which is a Winner-Takes-All type competitive network within N reference vectors, is an extended learning vector quantization whose distance measure is given by discriminant function of the Bayes decision. The extended LVQ is referred to learning Bayesian quantization (LBQ). LBQ organizes reference vectors which correspond to prototypes of input data through unsupervised learning. The second layer of the model includes N LBQs as its subnetworks. The a-th reference vector in the first layer connects to the a-th LBQ in the second layer. The a-th LBQ works as a secondary LBQ if and only if the a-th neuron excites. When input data consists of several clusters which include several subclusters, the presented model organizes the prototypes of the clusters as the reference vector in the first layer and those of the subclusters as the reference vector in the second layer hierarchically.