A parallel learning method of neural networks with feature extraction mechanism by autoencoder

カテゴリ: 部門大会

論文No: SS3-4

グループ名: 【C】平成30年電気学会電子・情報・システム部門大会プログラム

発行日: 2018/09/05

タイトル(英語): A parallel learning method of neural networks with feature extraction mechanism by autoencoder

著者名: 松井 央佑(千葉大学),岡本 卓(SENSY),小圷 成一(千葉大学),下馬場 朋禄(千葉大学),伊藤 智義(千葉大学)

著者名(英語): Osuke Matsui|Takashi Okamoto|Seiichi Koakutsu|Tomoyoshi Shimobaba|Tomoyoshi Ito

キーワード: ニューラルネットワーク|オートエンコーダ|デノイジングオートエンコーダデノイジングオートエンコーダ|neural network|autoencoder|denoising autoencoder

要約(日本語): In general, practical problems are complicated. A methodology of the machine learning by neural networks with a large number of layers is called deep learning. The superior capability of deep learning for practical problems is shown because of its higher level representation. However, learning problems cause a learning plateau.Previous researches have indicated the effectiveness of pre-training method. Pre-training obtains a good feature that is emblematic of a problem and initial value.This study proposes a pre-training method in parallel with fine-tuning by denoising autoencoder that is used corrupted input. The effectiveness of the proposed method is confirmed through computation experiments in which learning problems with corrupted data are used.

PDFファイルサイズ: 150 Kバイト