Paper Abstract and Keywords |
Presentation |
2015-01-23 09:50
Analysis of Minimum Classification Error Training using Bit-String-Based Genetic Algorithms Hiroto Togoe (Doshisha Univ.), Hideyuki Watanabe (NICT), Shigeru Katagiri (Doshisha Univ.), Xugang Lu, Chiori Hori (NICT), Miho Ohsaki (Doshisha Univ.) PRMU2014-100 MVE2014-62 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
Minimum Classification Error (MCE) training using gradient-descent-based loss minimization does not guarantee a global minimum of loss. To solve this problem, MCE training using a real-coded Genetic Algorithm (GA), which is considered suitable for global minimization, was investigated. However, its utility was not clearly demonstrated. In this paper, we newly apply another type of GA, i.e., a bit-string-based GA, to MCE training. From experiments, where such key features of MCE training as the smoothness of the classification error count loss were systematically controlled, we elaborate the nature of bit-string-based loss minimization for MCE training and show that GA-based minimization is not necessarily superior to handy gradient-descent-based minimization. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Minimum classification error training / Global loss minimization / Genetic algorithms / / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 114, no. 409, PRMU2014-100, pp. 171-176, Jan. 2015. |
Paper # |
PRMU2014-100 |
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 |
PRMU2014-100 MVE2014-62 |
|