Adaptive hamming net: A fast-learning ART 1 model without searching

Abstract

This paper introduces a neural network architecture called an adaptive Hamming net for learning of recognition categories. This model allows new prototypes to be added to an existing set of memorized prototypes without retraining the entire network. Under some model hypotheses, the functional behavior of the adaptive Hamming net is equivalent to that of a fast-learning ART 1 network, so some useful properties of ART I can be applied to the adaptive Hamming net. In addition, the proposed network finds the appropriate category more efficiently than ART 1: for the same input sequences, the adaptive Hamming net obtains the same recognition categories as ART 1 without any searching. The adaptive Hamming net not only reduces the training time of ART 1 but is also easier to implement. The adaptive Hamming net is limited to binary pattern clustering, but it can be extended to the case of analog input vectors by incorporating fuzzy logic techniques.