Simple structural information for generalization and rule extraction

Information called simple structural information is proposed. Simple structural information is composed of two types of information on the distribution of connections. By controlling structural information, we can freely control internal representations. We applied the method to two problems: a simple pattern classification and a language acquisition problem. In the pattern classification problem experimental results confirmed that the simple structural information can freely be controlled by the structural parameter. In addition different internal representations can be produced by controlling the simple structural information. In the language acquisition problem complex enough to test the performance, experimental results confirmed that generalization can be improved and suggested that explicit rules can be generated.