Generation and recall method for long-term memory data to suppress interference in ran
Tadahiko Murata, Yu Mizoguchi · 2005
A phenomenon called "interference" has been known as a problem in incremental learning of neural networks. That may happen when input-output relations trained by former training data are collapsed by new training data incremented later. In order to cope with "interference" problems, RAN-LTM (resource allocating network with long term memory) has been proposed. It has a long term memory (LTM) which stores input-output relations acquired by RAN. Then the relations in the LTM are recalled by a current training sample in order to suppress the interference in learning process. While RAN-LTM is effective to avoid the interference, there are many parameters to be set carefully in advance. In this paper, generation and retrieving methods of LTM data are proposed. The modified RAN-LTM is applied to two function approximation problems. Computational simulation results show that the modification is effective to improve the performance of RAN-LTM.