Explicit knowledge extraction in information-theoretic supervised multi-layered SOM

In this paper, we examine the effectiveness of SOM knowledge to train multi-layered neural networks. We have known that the SOM can produce very rich knowledge, used for visualization and class structure interpretation. It is expected that this SOM knowledge can be used for many different purposes in addition to visualization and interpretation. By using more flexible information-theoretic SOM, we examine the effectiveness of SOM knowledge for training multi-layered networks. We applied the method to the spam mail identification problem. We found that SOM knowledge greatly facilitated the learning of multi-layered networks and could be used to improve generalization performance.