Exploiting Learned Symmetries in Group Equivariant Convolutions

Attila Lengyel; Jan van Gemert

doi:10.1109/ICIP42928.2021.9506362

Exploiting Learned Symmetries in Group Equivariant Convolutions

Pattern Recognition and Bioinformatics

Research output: Chapter in Book/Conference proceedings/Edited volume › Conference contribution › Scientific › peer-review

3 Citations (Scopus)

Abstract

Group Equivariant Convolutions (GConvs) enable convolutional neural networks to be equivariant to various transformation groups, but at an additional parameter and compute cost. We investigate the filter parameters learned by GConvs and find certain conditions under which they become highly redundant. We show that GConvs can be efficiently decomposed into depthwise separable convolutions while preserving equivariance properties and demonstrate improved performance and data efficiency on two datasets. All code is publicly available at github.com/Attila94/SepGrouPy.

Original language	English
Title of host publication	2021 IEEE International Conference on Image Processing (ICIP)
Subtitle of host publication	Proceedings
Place of Publication	Piscataway
Publisher	IEEE
Pages	759-763
Number of pages	5
ISBN (Electronic)	978-1-6654-4115-5
ISBN (Print)	978-1-6654-3102-6
DOIs	https://doi.org/10.1109/ICIP42928.2021.9506362
Publication status	Published - 2021
Event	2021 IEEE International Conference on Image Processing (ICIP) - Virtual at Anchorage, United States Duration: 19 Sept 2021 → 22 Sept 2021

Conference

Conference	2021 IEEE International Conference on Image Processing (ICIP)
Country/Territory	United States
City	Virtual at Anchorage
Period	19/09/21 → 22/09/21

Keywords

group equivariant convolutions
depth-wise separable convolutions
efficient deep learning

Access to Document

10.1109/ICIP42928.2021.9506362

Cite this

@inproceedings{47818acbda1f43359ea936c56fbf087e,

title = "Exploiting Learned Symmetries in Group Equivariant Convolutions",

abstract = "Group Equivariant Convolutions (GConvs) enable convolutional neural networks to be equivariant to various transformation groups, but at an additional parameter and compute cost. We investigate the filter parameters learned by GConvs and find certain conditions under which they become highly redundant. We show that GConvs can be efficiently decomposed into depthwise separable convolutions while preserving equivariance properties and demonstrate improved performance and data efficiency on two datasets. All code is publicly available at github.com/Attila94/SepGrouPy.",

keywords = "group equivariant convolutions, depth-wise separable convolutions, efficient deep learning",

author = "Attila Lengyel and {van Gemert}, Jan",

year = "2021",

doi = "10.1109/ICIP42928.2021.9506362",

language = "English",

isbn = "978-1-6654-3102-6",

pages = "759--763",

booktitle = "2021 IEEE International Conference on Image Processing (ICIP)",

publisher = "IEEE",

address = "United States",

note = "2021 IEEE International Conference on Image Processing (ICIP) ; Conference date: 19-09-2021 Through 22-09-2021",

}

Lengyel, A & van Gemert, J 2021, Exploiting Learned Symmetries in Group Equivariant Convolutions. in 2021 IEEE International Conference on Image Processing (ICIP): Proceedings., 9506362, IEEE, Piscataway, pp. 759-763, 2021 IEEE International Conference on Image Processing (ICIP), Virtual at Anchorage, United States, 19/09/21. https://doi.org/10.1109/ICIP42928.2021.9506362

TY - GEN

T1 - Exploiting Learned Symmetries in Group Equivariant Convolutions

AU - Lengyel, Attila

AU - van Gemert, Jan

PY - 2021

Y1 - 2021

N2 - Group Equivariant Convolutions (GConvs) enable convolutional neural networks to be equivariant to various transformation groups, but at an additional parameter and compute cost. We investigate the filter parameters learned by GConvs and find certain conditions under which they become highly redundant. We show that GConvs can be efficiently decomposed into depthwise separable convolutions while preserving equivariance properties and demonstrate improved performance and data efficiency on two datasets. All code is publicly available at github.com/Attila94/SepGrouPy.

AB - Group Equivariant Convolutions (GConvs) enable convolutional neural networks to be equivariant to various transformation groups, but at an additional parameter and compute cost. We investigate the filter parameters learned by GConvs and find certain conditions under which they become highly redundant. We show that GConvs can be efficiently decomposed into depthwise separable convolutions while preserving equivariance properties and demonstrate improved performance and data efficiency on two datasets. All code is publicly available at github.com/Attila94/SepGrouPy.

KW - group equivariant convolutions

KW - depth-wise separable convolutions

KW - efficient deep learning

UR - http://www.scopus.com/inward/record.url?scp=85125590347&partnerID=8YFLogxK

U2 - 10.1109/ICIP42928.2021.9506362

DO - 10.1109/ICIP42928.2021.9506362

M3 - Conference contribution

SN - 978-1-6654-3102-6

SP - 759

EP - 763

BT - 2021 IEEE International Conference on Image Processing (ICIP)

PB - IEEE

CY - Piscataway

T2 - 2021 IEEE International Conference on Image Processing (ICIP)

Y2 - 19 September 2021 through 22 September 2021

ER -

Exploiting Learned Symmetries in Group Equivariant Convolutions

Abstract

Conference

Keywords

Access to Document

Other files and links

Fingerprint

On Color and Symmetries for Data Efficient Deep Learning

Code underlying the publication: Exploiting Learned Symmetries in Group Equivariant Convolutions

Cite this

Exploiting Learned Symmetries in Group Equivariant Convolutions

Abstract

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Research output

On Color and Symmetries for Data Efficient Deep Learning

Datasets

Code underlying the publication: Exploiting Learned Symmetries in Group Equivariant Convolutions

Cite this