An incremental learning method using weighted magnitude of interference
Nobuhiko Yamaguchi, Y. Minagawa, Koichiro Yamauchi, Naohiro Ishii · 2002
Almost all of neural network models need relearning of past learning patterns to memorize new patterns incrementally to avoid forgetting of past memories. This is due to each parameter of the neural network contributes to memorize several learning patterns. Therefore, the change in each parameter due to the learning interferes with the past memories. In this paper, we propose a new incremental learning method to omit the re-learning process to reduce the cost of the incremental learning. The method uses an objective function, which is to be minimized during the learning, called "weighted magnitude of interference." The function represents the change in output function of the neural network, which is caused by the modification of parameters and square error to the new patterns.