Self organizing networks with a split and merge algorithm

Abstract

Summary form only given, as follows. The authors present a novel learning algorithm for artificial neural networks based on the split and merge technique. The algorithm detects the similarity between the input patterns, and identifies the number of categories present in input samples. The algorithm is similar to the competitive learning algorithm; however, unlike in the competitive algorithm, the authors suggest two types of weights: long-term weights (LTWs) and short-term weights (STWs). The LTWs provide to the network the stability with respect to irrelevant input patterns, whereas the STWs provide the plasticity. The model with the split and merge algorithm has been developed and is used to categorize the pixels in the multispectral image based on the observed spectral signatures.<>