2012 Fiscal Year Final Research Report
STUDY ON BREAKING PERFORMANCE LIMITATINOS OF MINIMUM CLASSIFICATION ERROR TRAINING
| Project/Area Number |
22300064
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (B)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Perception information processing/Intelligent robotics
|
|---|
| Research Institution | Doshisha University |
|---|
Principal Investigator |
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
WATANABE Hideyuki (独)情報通信研究機構, ユニバーサルコミュニケーション研究所, 主任研究員 (40395091)
NAKAMURA Atsushi 日本電信電話(株)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, 主幹研究員 (50396206)
MATSUDA Shigeki (独)情報通信研究機構, ユニバーサルコミュニケーション研究所, 主任研究員 (20395007)
HORI Takaaki 日本電信電話(株)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)
WATANABE Shinji 日本電信電話(株)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, 研究員 (50396214)
|
|---|
| Project Period (FY) |
2010 – 2012
|
| Keywords | パターン認識 / 識別学習 / 最小分類誤り学習 / 計算論的学習理論 / 音声情報処理 |
|---|
| Research Abstract |
Aiming at further extending the discriminative power of the Minimum Classification Error (MCE) training, we focused on the analysis of the mechanism that the smoothness of smoothed classification error count loss increased training robustness to unseen pattern samples, and successfully achieved increase in the training’s discriminative power by developing an automatic optimization method for the smoothness and a novel Kernel MCE method. In addition, we developed a highly efficient discriminative training method, called Round-Robin Duel Discrimination (R2D2) method, for large-scale complex pattern classifiers such as large-vocabulary continuous speech recognizers, and successfully demonstrated its utility through systematic experimental evaluations.
|
