Explicit class structure produced by information-theoretic competitive and cooperative learning

In this paper, we propose a new type of information-theoretic method to realize self-organizing maps (SOM). In particular, the method can be used to produce a more explicit class structure than can the conventional SOM. The method is composed of a competitive network and a cooperative network that interact with each other. First, the cooperative network is used for training, and then the competitive network is introduced and interacts with the cooperative network. The influence of the competitive network is controlled by a parameter. When the parameter is one, the two networks are mutually enhanced; and when the parameter is increased, the competitive network´s influence is diminished. We applied the method to artificial data with three classes. We found that, first, with the cooperative network, the topographic errors could be decreased. Then, the introduction of the competitive network influenced the cooperative network. Better performance was obtained in terms of quantization and topographic errors. In addition, much sharper class boundaries were generated with the introduction of the competitive network.