マルチラベル・パターン認識機能を有する逐次マルチタスク学習モデル

カテゴリ: 部門大会

論文No: TC1-11

グループ名: 【C】平成26年電気学会電子・情報・システム部門大会講演論文集

発行日: 2014/09/03

タイトル(英語): A Sequential Multi-task Learning Model with Multi-label Pattern Recognition Function

著者名: 樋口 大輔(神戸大学),小澤 誠一(神戸大学)

著者名(英語): Daisuke Higuchi(Kobe University),Seiichi Ozawa(Kobe University)

キーワード: マルチタスク学習|マルチラベルパターン認識|ニューラルネット|追加学習|Multi-task Learning|Multi-label Pattern Recognition|Neural Networks|Incremental Learning

要約(日本語): In this paper, we propose an incremental neural network model for a general class of sequential multi-task classification problems where a training data may have multiple class labels. To handle this type of classification problems, the proposed model consists of a three-layer feedforward neural network with long-term/short-term memories, and it has the following functions: one-pass incremental learning, task allocation, handling multi-label data, task consolidation, and knowledge transfer. In the experiments, we evaluate the proposed model for various kinds of data sets. The experimental results demonstrate that the proposed model has good performance in both classification and task categorization even if the task information is not always given.

PDFファイルサイズ: 186 Kバイト