Exploring Inter-Channel Correlation for Diversity-preserved Knowledge Distillation

Liu, L; Huang, Q; Lin, S; Xie, H; Wang, B; Chang, X; Liang, X

Exploring Inter-Channel Correlation for Diversity-preserved Knowledge Distillation

Liu, L Huang, Q Lin, S Xie, H Wang, B Chang, X

Liang, X

Permalink

Publisher:: IEEE
Publication Type:: Conference Proceeding
Citation:: Proceedings of the IEEE International Conference on Computer Vision, 2021, 00, pp. 8251-8260
Issue Date:: 2021-01-01

Closed Access

	Filename	Description	Size
	KD_ICCV2021 (1).pdf	Submitted version	6.36 MB	Adobe PDF	View/Open

Copyright Clearance Process

Recently Added
In Progress
Closed Access

This item is closed access and not available.

Full metadata record

Field	Value	Language
dc.contributor.author	Liu, L
dc.contributor.author	Huang, Q
dc.contributor.author	Lin, S
dc.contributor.author	Xie, H
dc.contributor.author	Wang, B
dc.contributor.author	Chang, X https://orcid.org/0000-0002-7778-8807
dc.contributor.author	Liang, X
dc.date	2021-10-11
dc.date.accessioned	2023-03-31T09:26:34Z
dc.date.available	2023-03-31T09:26:34Z
dc.date.issued	2021-01-01
dc.identifier.citation	Proceedings of the IEEE International Conference on Computer Vision, 2021, 00, pp. 8251-8260
dc.identifier.isbn	9781665428125
dc.identifier.issn	1550-5499
dc.identifier.uri	http://hdl.handle.net/10453/168960
dc.description.abstract	Knowledge Distillation has shown very promising ability in transferring learned representation from the larger model (teacher) to the smaller one (student). Despite many efforts, prior methods ignore the important role of retaining inter-channel correlation of features, leading to the lack of capturing intrinsic distribution of the feature space and sufficient diversity properties of features in the teacher network. To solve the issue, we propose the novel Inter-Channel Correlation for Knowledge Distillation (ICKD), with which the diversity and homology of the feature space of the student network can align with that of the teacher network. The correlation between these two channels is interpreted as diversity if they are irrelevant to each other, otherwise homology. Then the student is required to mimic the correlation within its own embedding space. In addition, we introduce the grid-level inter-channel correlation, making it capable of dense prediction tasks. Extensive experiments on two vision tasks, including ImageNet classification and Pascal VOC segmentation, demonstrate the superiority of our ICKD, which consistently outperforms many existing methods, advancing the state-of-the-art in the fields of Knowledge Distillation. To our knowledge, we are the first method based on knowledge distillation boosts ResNet18 beyond 72% Top-1 accuracy on ImageNet classification. Code is available at: https://github.com/ADLab-AutoDrive/ICKD.
dc.language	en
dc.publisher	IEEE
dc.relation	http://purl.org/au-research/grants/arc/DE190100626
dc.relation.ispartof	Proceedings of the IEEE International Conference on Computer Vision
dc.relation.ispartof	18th IEEE/CVF International Conference on Computer Vision (ICCV)
dc.relation.isbasedon	10.1109/ICCV48922.2021.00816
dc.rights	info:eu-repo/semantics/closedAccess
dc.title	Exploring Inter-Channel Correlation for Diversity-preserved Knowledge Distillation
dc.type	Conference Proceeding
utslib.citation.volume	00
utslib.location.activity	ELECTR NETWORK
pubs.organisational-group	/University of Technology Sydney
pubs.organisational-group	/University of Technology Sydney/Faculty of Engineering and Information Technology
pubs.organisational-group	/University of Technology Sydney/Strength - AAII - Australian Artificial Intelligence Institute
utslib.copyright.status	closed_access	*
dc.date.updated	2023-03-31T09:26:30Z
pubs.finish-date	2021-10-17
pubs.publication-status	Published
pubs.start-date	2021-10-11
pubs.volume	00

Abstract:

Knowledge Distillation has shown very promising ability in transferring learned representation from the larger model (teacher) to the smaller one (student). Despite many efforts, prior methods ignore the important role of retaining inter-channel correlation of features, leading to the lack of capturing intrinsic distribution of the feature space and sufficient diversity properties of features in the teacher network. To solve the issue, we propose the novel Inter-Channel Correlation for Knowledge Distillation (ICKD), with which the diversity and homology of the feature space of the student network can align with that of the teacher network. The correlation between these two channels is interpreted as diversity if they are irrelevant to each other, otherwise homology. Then the student is required to mimic the correlation within its own embedding space. In addition, we introduce the grid-level inter-channel correlation, making it capable of dense prediction tasks. Extensive experiments on two vision tasks, including ImageNet classification and Pascal VOC segmentation, demonstrate the superiority of our ICKD, which consistently outperforms many existing methods, advancing the state-of-the-art in the fields of Knowledge Distillation. To our knowledge, we are the first method based on knowledge distillation boosts ResNet18 beyond 72% Top-1 accuracy on ImageNet classification. Code is available at: https://github.com/ADLab-AutoDrive/ICKD.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/168960