Presentation | 2001/10/11 Unsupervised learning for single hypothesis problem via mixtures of modified quadratic discriminant functions Tsuyoshi KATO, Shinichiro OMACHI, Hirotomo ASO, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | This paper shows that the performance of the single hypothesis scheme-it classifies unknown patterns to the specified class called target class-can be improved by estimating the probabilistic distribution of nontarget class using additional unlabeled training patterns that are available at no extra cost. Furthermore, our method achives comparable recognition accuracy to the supervised classifier, yet requires no labeled nontarget patterns for training. We introduce an algorithm for learning nontargets's pdf.from unlabeled patterns based on Expectation-Maximization and the proposed stochastic model, the mixture of modified quadratic discriminant function(MQDF). The MQDF is well-known as one of the state-of-the-art classifiers, however, it suffers from the weakness for multi-modal distribution. The mixture of MQDF overcomes this problem, yet which reteins the excellent classification performance. We demonstrate the practicality and accuracy of the mixture of MQDF estimated by the proposed unsupervised learning using the artificial and the real world databases. Forthermore, the proposed algorithm is also tested on the letter-image identification problem. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | single hypothesis problem / mixture of modified quadratic discriminant funcions / unsupervised learning / Expectation-Maximization algorithm / letter-image identifier |
Paper # | PRMU2001-103,NC2001-53 |
Date of Issue |
Conference Information | |
Committee | PRMU |
---|---|
Conference Date | 2001/10/11(1days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | |
Chair | |
Vice Chair | |
Secretary | |
Assistant |
Paper Information | |
Registration To | Pattern Recognition and Media Understanding (PRMU) |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Unsupervised learning for single hypothesis problem via mixtures of modified quadratic discriminant functions |
Sub Title (in English) | |
Keyword(1) | single hypothesis problem |
Keyword(2) | mixture of modified quadratic discriminant funcions |
Keyword(3) | unsupervised learning |
Keyword(4) | Expectation-Maximization algorithm |
Keyword(5) | letter-image identifier |
1st Author's Name | Tsuyoshi KATO |
1st Author's Affiliation | Graduate School of Engineering, Tohoku University() |
2nd Author's Name | Shinichiro OMACHI |
2nd Author's Affiliation | Graduate School of Engineering, Tohoku University |
3rd Author's Name | Hirotomo ASO |
3rd Author's Affiliation | Graduate School of Engineering, Tohoku University |
Date | 2001/10/11 |
Paper # | PRMU2001-103,NC2001-53 |
Volume (vol) | vol.101 |
Number (no) | 362 |
Page | pp.pp.- |
#Pages | 8 |
Date of Issue |