Information Maximization for Variable Selection in Competitive Learning

R. Kamimura, K. Aoyama, and Y. Asamizu (Japan)


Competitive learning, self-organizing maps, mutual information, information maximization, enhanced information, variable selection


In this paper, we introduce an information maximization method to describe variable selection in competitive learning, in particular, SOM(self-organizing maps) by introducing information-theoretic importance of input variables. In this method, information content concerning the importance of input variables is supposed to be maximized so as to condense information on input variables as much as possible. The importance of input variables is determined by enhanced information. The enhanced information is mutual information between input patterns and competitive units with enhanced attention to given input variables. Then, connection weights are determined so as to maximize information based upon this enhanced information about input variables. We applied the method to two well-known data from machine-learning database, namely, the wine and the voting attitude data. Experimental results conīŦrmed that by applying the method to two problems, the performance of the conventional SOM can be improved in terms of quantization and topographic errors.

Important Links:

Go Back