2017 Fiscal Year Final Research Report

Minimum Classification Error Criterion-based Development of Highly Discriminative Feature Space Associated with Optimal Class Boundary Search Methods

Research Project

PDF

Project/Area Number	26280063
Research Category	Grant-in-Aid for Scientific Research (B)
Allocation Type	Partial Multi-year Fund
Section	一般
Research Field	Perceptual information processing
Research Institution	Doshisha University
Principal Investigator	Katagiri Shigeru 同志社大学, 理工学部, 教授 (40396114)
Co-Investigator(Kenkyū-buntansha)	渡辺秀行株式会社国際電気通信基礎技術研究所, その他部局等, 研究員 (40395091) 中村篤名古屋市立大学, 大学院システム自然科学研究科, 教授 (50396206) Delcroix Marc 日本電信電話株式会社NTTコミュニケーション科学基礎研究所, 協創情報研究部, 主任研究員 (70793339) 小川厚徳日本電信電話株式会社NTTコミュニケーション科学基礎研究所, メディア情報研究部, 主任研究員 (90527516) 吉岡拓也日本電信電話株式会社NTTコミュニケーション科学基礎研究所, メディア情報研究部, 研究主任 (40466404) 堀貴明日本電信電話株式会社NTTコミュニケーション科学基礎研究所, メディア情報研究部, 主任研究員 (20396211)
Project Period (FY)	2014-04-01 – 2018-03-31
Keywords	パターン認識 / 識別学習 / 最小分類誤り学習 / カーネル法 / ニューラルネットワーク / 音声認識 / 機械学習
Outline of Final Research Achievements	Aiming at the development of highly discriminative feature space, of which corresponding classification error probability is as small as possible, we developed the following new technologies: a Dynamic-Time-Warping (DTW)-based geometric margin for variable-length patterns, Large Geometric Margin Minimum Classification Error training using the DTW-based geometric margin, a compact kernel classifier using Kernel Minimum Classification Error training, speaker and environment adaptation methods for deep-neural-network-based speech recognizers using Speaker Adaptive Training and auxiliary neural network, and fast search methods for large scale speech recognizers. In addition, we opened a new venue for a new pattern recognizer training method that does not require hyper-parameters but is based on Bayes boundary-ness, which is defined using the ambiguity in classification decision around estimated class boundaries.
Free Research Field	人間情報学（知覚情報処理），知能情報学