Separation and unification of individuality and collectivity and its application to explicit class structure in self-organizing maps

Abstract

In this paper, we propose a new type of learning method in which individuality and collectivity are separated and unified to control the characteristics of neurons. This unification is expected to enhance the characteristics shared by individual and collective outputs, while the characteristics specific to them are weakened. We applied the method to self-organizing maps to demonstrate the utility of unification. In self-organizing maps, the introduction of unification has the effect of controlling cooperation among neurons. Experimental results on the glass identification problem from the machine learning database showed that explicit class boundaries could be obtained by introducing the unification. © 2012 Springer-Verlag.