Paper Abstract and Keywords |
Presentation |
2009-07-17 13:50
Geometric Margin Control for Minimum Error Classification Kouta Yamada, Shigeru Katagiri (Doshisha Univ.), Erik McDermott (NTT), Hideyuki Watanabe (NICT), Atsushi Nakamura, Shinji Watanabe (NTT), Miho Ohsaki (Doshisha Univ.) SP2009-43 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
The recent dramatic growth of computation power and data availability has increased research interests in discriminative training methods for pattern classifier design, especially the Minimum Classification Error (MCE) training method that directly aims to minimize misclassification rate, and the Support Vector Machine (SVM) method that aims to improve the training robustness. In this paper, we present a noble, highly robust version of MCE, which estimates more accurately than before the true minimum classification error condition with breaking the curse of over-fitting to training samples, by newly formulating a geometric margin of a practical discriminant funciton that uses distance or probability measures. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Minimum Classification Error training / geometric margin / functional margin / Support Vector Machine / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 109, no. 139, SP2009-43, pp. 13-18, July 2009. |
Paper # |
SP2009-43 |
Date of Issue |
2009-07-10 (SP) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
Download PDF |
SP2009-43 |
|