Paper Abstract and Keywords |
Presentation |
2019-06-18 14:45
Theoretical Analysis of the Fixup Initialization for Fast Convergence and High Generalization Ability Yasutaka Furusho, Kazushi Ikeda (NAIST) NC2019-19 IBISML2019-17 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
The Fixup initialization is a new initialization method of ResNet for a fast convergence with a high learning rate of SGD and a high generalization ability. However, the reasons for its high performance are not clear.
Both the learning rate and the generalization ability are affected by the loss landscape at initialization, that is, the maximum eigenvalue of the Hessian matrix. Thus, we calculated the maximum eigenvalue of the ResNet and found that the maximum eigenvalue with the Fixup initialization has at most the square root order with respect to the depth of the ResNet while that with the He initialization has the exponential order. This small eigenvalue leads to a fast convergence with a high learning rate of SGD and a high generalization ability. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Deep neural network / ResNet / Initialization / Fixup / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 119, no. 89, IBISML2019-17, pp. 109-114, June 2019. |
Paper # |
IBISML2019-17 |
Date of Issue |
2019-06-10 (NC, IBISML) |
ISSN |
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 |
NC2019-19 IBISML2019-17 |
Conference Information |
Committee |
NC IBISML IPSJ-MPS IPSJ-BIO |
Conference Date |
2019-06-17 - 2019-06-19 |
Place (in Japanese) |
(See Japanese page) |
Place (in English) |
Okinawa Institute of Science and Technology |
Topics (in Japanese) |
(See Japanese page) |
Topics (in English) |
Neurocomputing, Machine Learning Approach to Biodata Mining, and General |
Paper Information |
Registration To |
IBISML |
Conference Code |
2019-06-NC-IBISML-MPS-BIO |
Language |
Japanese |
Title (in Japanese) |
(See Japanese page) |
Sub Title (in Japanese) |
(See Japanese page) |
Title (in English) |
Theoretical Analysis of the Fixup Initialization for Fast Convergence and High Generalization Ability |
Sub Title (in English) |
|
Keyword(1) |
Deep neural network |
Keyword(2) |
ResNet |
Keyword(3) |
Initialization |
Keyword(4) |
Fixup |
Keyword(5) |
|
Keyword(6) |
|
Keyword(7) |
|
Keyword(8) |
|
1st Author's Name |
Yasutaka Furusho |
1st Author's Affiliation |
Nara Institute of Science and Technology (NAIST) |
2nd Author's Name |
Kazushi Ikeda |
2nd Author's Affiliation |
Nara Institute of Science and Technology (NAIST) |
3rd Author's Name |
|
3rd Author's Affiliation |
() |
4th Author's Name |
|
4th Author's Affiliation |
() |
5th Author's Name |
|
5th Author's Affiliation |
() |
6th Author's Name |
|
6th Author's Affiliation |
() |
7th Author's Name |
|
7th Author's Affiliation |
() |
8th Author's Name |
|
8th Author's Affiliation |
() |
9th Author's Name |
|
9th Author's Affiliation |
() |
10th Author's Name |
|
10th Author's Affiliation |
() |
11th Author's Name |
|
11th Author's Affiliation |
() |
12th Author's Name |
|
12th Author's Affiliation |
() |
13th Author's Name |
|
13th Author's Affiliation |
() |
14th Author's Name |
|
14th Author's Affiliation |
() |
15th Author's Name |
|
15th Author's Affiliation |
() |
16th Author's Name |
|
16th Author's Affiliation |
() |
17th Author's Name |
|
17th Author's Affiliation |
() |
18th Author's Name |
|
18th Author's Affiliation |
() |
19th Author's Name |
|
19th Author's Affiliation |
() |
20th Author's Name |
|
20th Author's Affiliation |
() |
Speaker |
Author-1 |
Date Time |
2019-06-18 14:45:00 |
Presentation Time |
25 minutes |
Registration for |
IBISML |
Paper # |
NC2019-19, IBISML2019-17 |
Volume (vol) |
vol.119 |
Number (no) |
no.88(NC), no.89(IBISML) |
Page |
pp.87-92(NC), pp.109-114(IBISML) |
#Pages |
6 |
Date of Issue |
2019-06-10 (NC, IBISML) |
|