Paper Abstract and Keywords |
Presentation |
2022-12-16 15:10
[Short Paper]
Cosine Similarity Based Attention on a Hypersphere for Vision Transformers Jungdae Lee, Rei Kawakami, Nakamasa Inoue (Tokyo Tech) PRMU2022-52 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
The success of Vision Transformers in computer vision is usually attributed to its distinct structure of self-attention architecture. However, it has been pointed out by a number of researchers that self-attention in Vision Transformer can be replaced with more general structures that involve mixing intermediate tokens. In this paper, we propose an architecture that generalizes self-attention block of Vision Transformer as a loss function combined with a classifier. On top of this, we present an instance of the architecture that employs a combination of logistic regression losses in conjunction with a linear classifier, which relies on cosine similarity instead of dot product. In the experiment section of the paper, we compared the proposed model with DeiT, an effective light-weight Transformer, on classification accuracy for CIFAR10 and CIFAR100 dataset without external training data. As a result, we have discovered that the proposed model performs better than DeiT under certain settings. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Vision Transformer / attention / loss function / L2 normalization / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 122, no. 314, PRMU2022-52, pp. 106-109, Dec. 2022. |
Paper # |
PRMU2022-52 |
Date of Issue |
2022-12-08 (PRMU) |
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 |
PRMU2022-52 |
Conference Information |
Committee |
PRMU |
Conference Date |
2022-12-15 - 2022-12-16 |
Place (in Japanese) |
(See Japanese page) |
Place (in English) |
Toyama International Conference Center |
Topics (in Japanese) |
(See Japanese page) |
Topics (in English) |
|
Paper Information |
Registration To |
PRMU |
Conference Code |
2022-12-PRMU |
Language |
English (Japanese title is available) |
Title (in Japanese) |
(See Japanese page) |
Sub Title (in Japanese) |
(See Japanese page) |
Title (in English) |
Cosine Similarity Based Attention on a Hypersphere for Vision Transformers |
Sub Title (in English) |
|
Keyword(1) |
Vision Transformer |
Keyword(2) |
attention |
Keyword(3) |
loss function |
Keyword(4) |
L2 normalization |
Keyword(5) |
|
Keyword(6) |
|
Keyword(7) |
|
Keyword(8) |
|
1st Author's Name |
Jungdae Lee |
1st Author's Affiliation |
Tokyo Institute of Technology (Tokyo Tech) |
2nd Author's Name |
Rei Kawakami |
2nd Author's Affiliation |
Tokyo Institute of Technology (Tokyo Tech) |
3rd Author's Name |
Nakamasa Inoue |
3rd Author's Affiliation |
Tokyo Institute of Technology (Tokyo Tech) |
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 |
2022-12-16 15:10:00 |
Presentation Time |
10 minutes |
Registration for |
PRMU |
Paper # |
PRMU2022-52 |
Volume (vol) |
vol.122 |
Number (no) |
no.314 |
Page |
pp.106-109 |
#Pages |
4 |
Date of Issue |
2022-12-08 (PRMU) |
|