Paper Abstract and Keywords |
Presentation |
2014-01-23 09:30
Minimum Classification Error Training with Automatic Determination of Loss Smoothness Common to All Classes Kensuke Ota (Doshisha Univ.), Hideyuki Watanabe (NICT), Shigeru Katagiri, Miho Ohsaki (Doshisha Univ.), Shigeki Matsuda, Chiori Hori (NICT) PRMU2013-91 MVE2013-32 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
The smoothness of the smooth classification error count loss used in the Minimum Classification Error (MCE) training has an effect of increasing training robustness to unseen samples. Therefore, an appropriate determination of the smoothness is obviously needed. Recently, to meet this necessity, a method using the Parzen-estimation-based MCE formalization was proposed for automatically setting the smoothness, and its effectiveness was demonstrated. However, this method sets the smoothness in the class-by-class mode, and it has a potential risk of causing over-fitting to training samples. In this paper, we propose a new method for automatically finding an appropriate value of the smoothness that is set to all of the classes, and demonstrate its usefulness. From evaluation experiments, we show that the proposed method works more stably and more effectively under various classifier conditions than its counterpart, preceding automatic determination method. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Minimum classification error training / Parzen estimation / / / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 113, no. 402, PRMU2013-91, pp. 1-6, Jan. 2014. |
Paper # |
PRMU2013-91 |
Date of Issue |
2014-01-16 (PRMU, MVE) |
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 |
PRMU2013-91 MVE2013-32 |
|