Contradiction Resolution and its Application to Self-Organizing Maps

R. Kamimura (Japan)

Keywords

Competitive learning, selforganizing maps, contradiction resolution, free energy

Abstract

In this paper, we propose a new learning method called contradiction resolution, which is applied to the generation of self-organizing maps. The method is composed of two steps, namely, contradiction perception and resolution. First, contradiction or difference is perceived among different terms, states or components in a neural network. Once the contradiction is perceived, contradiction resolution can be used to reduce the contradiction as much as possible. One of the main differences from other learning methods is that contradiction resolution does not necessarily solve the contradiction, but effects a compromise between contradictory terms as much as possible. Thus, the method can be applied to learning methods where explicit targets cannot be identified, for example, unsupervised learning. We applied the method to the generation of self-organizing maps by using two problems. In both problems, in terms of the quantization and topological errors, conflict resolution outperformed the conventional SOM, but it did not necessarily give better results in terms of two topology preservation measures such as trustworthiness and continuity.

Important Links:



Go Back