A modular type network for incremental learning

A modular type network, which consists of the same number of modules as that of classes, is an efficient way of using layered neural networks for multiclass classification problems. In the case where training patterns of a new class are added after training the initial pattern set is completed we usually need to re-train both old modules and the new module which corresponds to the new class. In this case, to use all of the training patterns for re-training is not an efficient way. But any efficient neural networks for this type of incremental learning with modular type networks have not been presented. In this paper, we propose a new modular type network and its learning algorithm for the incremental learning that don't damage classification capability acquired already and select input patterns to learn adaptively when the network is re-trained.