Pseudo-network Growing for Gradual Interpretation of Input Patterns

Abstract

In this paper, we propose a new information-theoretic method to interpret competitive learning. The method is called ”pseudo-network growing,” because a network re-grows gradually after learning, taking into account the importance of components. In particular, we try to apply the method to clarify the class structure of self-organizing maps. First, the importance of input units is computed, and then input units are gradually added, according to their importance. We can expect that the corresponding number of competitive units will be gradually increased, showing the main characteristics of network configurations and input patterns. We applied the method to the well-known Senate data with two distinct classes. By using the conventional SOM, explicit class boundaries could not be obtained, due to the inappropriate map size imposed in the experiment. However, with the pseudo-network growing, a clear boundary could be observed in the first growing stage, and gradually the detailed class structure could be reproduced.