Information acquisition performance by supervised information-theoretic self-organizing maps

In this paper, we propose a new type of supervised multi-layered self-organizing map and examine to what extent information content in multi-layered networks can be increased. We have so far introduced the information-theoretic SOM in a single layer for increasing information content. However, we have found some cases where information content cannot be increased by single-layer networks. We used the multi-layered network and we found that mutual information tended to increase even for higher layers. The corresponding U-matrices showed clearer class structure even for higher layers. Then, we applied the method to the improvement of prediction performance. The prediction performance could be improved when the number of layers was appropriately chosen.