In this paper, we propose a new algorithm to produce explicit self-organizing maps. We suppose that individual neurons try to assimilate or imitate the collective behaviors of neighboring neurons as much as possible. The side effect of this assimilation consists in the generation of self-organizing maps. In the usual selforganizing maps' formulation, neurons are more closely related to each other as distance between neurons is closer. Thus, a collective behavior is that neighboring neurons behave quite similarly to each other. In the actual formulation and to obtain update rules, we used Gaussian mixture models with the conventional EM algorithm. Though the mixture models are not exact, experimental results on an artificial data and the Iris problem showed that clearer feature maps could be obtained, and even when the network size became large, the clear maps remained the same.
展开▼
