Incremental Learning Method of GRBF with Recalling of Interfered Patterns

This paper proposes a low-cost incremental learning method of Generalized Radial Basis Function (GRBF) for a Case Based Reasoning (CBR) system. A CBR system is one type of reasoning system that uses past cases for solving new problems. To realize the reasoning, the system has to search a past case which is similar to the new problem from a case database. If such case is found, the system has to adapt it so as to t the new problem. The adapted case is the presented solution. If the solution is not good, it is revised by an expert who gives a correct solution and stores it into the case database as a new case. If a system uses a neural network as its case database, the searching process and the adaptation process are not needed. The system only has to present the new problem to the neural network. Then, an appropriate solution appears in the output layer of the neural network. The neural network has to learn new cases completely when the case appears. However, if the network learns the new cases only by referring to them, the network probably forget old memorized cases. A certain way to avoid the lost of memory is learning the new cases with all memorized cases. It needs, however, a high computational power. To solve this problem, we propose an Incremental Learning method with Recalling Interfered patterns (ILRI) for Generalized Radial Basis Function (GRBF) [T. Poggio et al.,1990]. In ILRI, the GRBF learns new cases with relearning of the few recalled past cases that are interfered with the incremental learning.