Paper Abstract and Keywords |
Presentation |
2015-01-23 10:15
Relation between Data Grouping and Robustness to Unseen Data in Large Geometric Margin Minimum Classification Error Training Hiroyuki Shiraishi (Doshisha Univ), Hideyuki Watanabe (NICT), Shigeru Katagiri (Doshisha Univ), Xugang Lu, Chiori Hori (NICT), Miho Ohsaki (Doshisha Univ) PRMU2014-101 MVE2014-63 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
To develop a pattern classifier that is robust to unseen pattern samples, classifier parameters have been conventionally trained using both training and validation sample sets. However, there are no clear criteria for dividing the samples in hand into training and validation sets. In addition, such grouping decreases the number of samples for both training and validation, often lowering robustness to unseen samples. To solve this problem, we elaborate in this paper the nature of an approach that aims, without validation samples, for high robustness only with Large Geometric Margin Minimum Classification Error training over training samples. From experiments using several different sizes of training/validation sample sets, we clarify the advantages and disadvantages of the conventional approach using validation samples and show the potential utility of our proposed large-geometric-margin-based approach. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Pattern recognition / Minimum classification error training / Geometric margin / Data grouping for training / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 114, no. 409, PRMU2014-101, pp. 177-182, Jan. 2015. |
Paper # |
PRMU2014-101 |
Date of Issue |
2015-01-15 (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 |
PRMU2014-101 MVE2014-63 |
|