最適汎化能力をもった神経回路網の追加学習

概要

論文の詳細を見る
Learning in a Neural Network(NN) is considered as an inverse problem from the functional analytical point of view. These kind of approaches have been proved to improve generalizing ability because of their focus on the original function space rather than the sampled space. In earlier research work, learning has been discussed under the framework of obtaining an optimal learning operator from a given set of training data. However, there is a need to look at cases when we have an optimally trained NN and there are additions to the training data. Relearning using the entire training set all over again is a possible but not very efficient solution. These kind of situations also arise in 'active learning' in which the set of training data can be provided only one after the other and also uses the intermediate results hr deciding the next best sample. In this paper, we present an efficient method of incrementally carrying out learning while maintaining the optimal generalization ability at every stage and also clarify the relationship between the newly learned function and the previous function, opening avenues for work into selection of optimal training data.
社団法人電子情報通信学会の論文
1996-03-11