神経回路網における最適モデル選択---シミュレーションを用いた基礎的研究

カテゴリ: 部門大会

論文No: TC4-23

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): How Does the Neural Network Choose the Optimal Model?---A Basic Research Using the Computer Simulation

著者名: 磯村 拓哉(東京大学,日本学術振興会),小谷 潔(東京大学),神保 泰彦(東京大学)

著者名(英語): Takuya Isomura(The University of Tokyo,JSPS),Kiyoshi Kotani(The University of Tokyo),Yasuhiko Jimbo(The University of Tokyo)

キーワード: モデル選択|ニューラルネットワーク|独立成分分析|コンピュータシミュレーション|勾配学習法|model selection|neural network|independent component analysis|computer simulation|gradient learning method

要約(日本語): Unsupervised model selection is a new area of theoretical neuroscience. Unsupervised learning of variables and parameters has been frequently studied under the assumption that we have already known the model structures; however, how the neural network learns the optimal model structure has rarely been investigated. Here, we proposed a new semi-gradient algorithm of model selection performed by neural networks. Using the simulation of a 2-layer independent component analysis model, we evaluated the accuracy of estimating the model structure and the parameters of the generation model, and confirmed that the estimation error gradually altered as the change in the number of synaptic connections, suggesting that our algorithm can choose the optimal model according to the gradient scheme.

PDFファイルサイズ: 737 Kバイト