Unified Information Theoretic Approach to Competitive Learning

R. Kamimura, T. Kamimura, O. Uchida, and S. Nakanishi (Japan)

Keywords

Abstract

In this paper, we attempt to show that the unification of infor mation maximization and minimization can describe and explain many aspects of neural computing. Especially, we deal with com petitive learning for demonstrating this hypothesis and show that competition in competitive learning can be realized by unifying information maximization and minimization in neural networks. In competitive learning, information in input patterns should be increased as much as possible. This process is naturally trans lated into information maximization. In addition, we can see that all competitive units must be equally or impartially used, mean ing that a different competitive unit responds to a different input pattern. This equal or impartial use can be translated as a pro cess of information minimization. Thus, competitive learning can be realized by unified approach to neural computing. Compet itive learning is one of the good examples to show that unified approach can reveal many aspects of neural learning.

Important Links:



Go Back


IASTED
Rotating Call For Paper Image