Knowledge Transfer via Dense Cross-Layer Mutual-Distillation
Knowledge Distillation (KD) based methods adopt the one-way Knowledge Transfer (KT) scheme in which training a lower-capacity student network is guided by a pre-trained high-capacity teacher network. Recently, Deep Mutual Learning (DML) presented a two-way KT strategy, showing that the student netwo...
Main Authors: | , |
---|---|
Format: | Article in Journal/Newspaper |
Language: | unknown |
Published: |
arXiv
2020
|
Subjects: | |
Online Access: | https://dx.doi.org/10.48550/arxiv.2008.07816 https://arxiv.org/abs/2008.07816 |
id |
ftdatacite:10.48550/arxiv.2008.07816 |
---|---|
record_format |
openpolar |
spelling |
ftdatacite:10.48550/arxiv.2008.07816 2023-05-15T16:01:58+02:00 Knowledge Transfer via Dense Cross-Layer Mutual-Distillation Yao, Anbang Sun, Dawei 2020 https://dx.doi.org/10.48550/arxiv.2008.07816 https://arxiv.org/abs/2008.07816 unknown arXiv arXiv.org perpetual, non-exclusive license http://arxiv.org/licenses/nonexclusive-distrib/1.0/ Computer Vision and Pattern Recognition cs.CV Machine Learning cs.LG FOS Computer and information sciences Article CreativeWork article Preprint 2020 ftdatacite https://doi.org/10.48550/arxiv.2008.07816 2022-03-10T15:43:30Z Knowledge Distillation (KD) based methods adopt the one-way Knowledge Transfer (KT) scheme in which training a lower-capacity student network is guided by a pre-trained high-capacity teacher network. Recently, Deep Mutual Learning (DML) presented a two-way KT strategy, showing that the student network can be also helpful to improve the teacher network. In this paper, we propose Dense Cross-layer Mutual-distillation (DCM), an improved two-way KT method in which the teacher and student networks are trained collaboratively from scratch. To augment knowledge representation learning, well-designed auxiliary classifiers are added to certain hidden layers of both teacher and student networks. To boost KT performance, we introduce dense bidirectional KD operations between the layers appended with classifiers. After training, all auxiliary classifiers are discarded, and thus there are no extra parameters introduced to final models. We test our method on a variety of KT tasks, showing its superiorities over related methods. Code is available at https://github.com/sundw2014/DCM : Accepted by ECCV 2020. The code is available at https://github.com/sundw2014/DCM, which is based on the implementation of our DKS work https://github.com/sundw2014/DKS Article in Journal/Newspaper DML DataCite Metadata Store (German National Library of Science and Technology) |
institution |
Open Polar |
collection |
DataCite Metadata Store (German National Library of Science and Technology) |
op_collection_id |
ftdatacite |
language |
unknown |
topic |
Computer Vision and Pattern Recognition cs.CV Machine Learning cs.LG FOS Computer and information sciences |
spellingShingle |
Computer Vision and Pattern Recognition cs.CV Machine Learning cs.LG FOS Computer and information sciences Yao, Anbang Sun, Dawei Knowledge Transfer via Dense Cross-Layer Mutual-Distillation |
topic_facet |
Computer Vision and Pattern Recognition cs.CV Machine Learning cs.LG FOS Computer and information sciences |
description |
Knowledge Distillation (KD) based methods adopt the one-way Knowledge Transfer (KT) scheme in which training a lower-capacity student network is guided by a pre-trained high-capacity teacher network. Recently, Deep Mutual Learning (DML) presented a two-way KT strategy, showing that the student network can be also helpful to improve the teacher network. In this paper, we propose Dense Cross-layer Mutual-distillation (DCM), an improved two-way KT method in which the teacher and student networks are trained collaboratively from scratch. To augment knowledge representation learning, well-designed auxiliary classifiers are added to certain hidden layers of both teacher and student networks. To boost KT performance, we introduce dense bidirectional KD operations between the layers appended with classifiers. After training, all auxiliary classifiers are discarded, and thus there are no extra parameters introduced to final models. We test our method on a variety of KT tasks, showing its superiorities over related methods. Code is available at https://github.com/sundw2014/DCM : Accepted by ECCV 2020. The code is available at https://github.com/sundw2014/DCM, which is based on the implementation of our DKS work https://github.com/sundw2014/DKS |
format |
Article in Journal/Newspaper |
author |
Yao, Anbang Sun, Dawei |
author_facet |
Yao, Anbang Sun, Dawei |
author_sort |
Yao, Anbang |
title |
Knowledge Transfer via Dense Cross-Layer Mutual-Distillation |
title_short |
Knowledge Transfer via Dense Cross-Layer Mutual-Distillation |
title_full |
Knowledge Transfer via Dense Cross-Layer Mutual-Distillation |
title_fullStr |
Knowledge Transfer via Dense Cross-Layer Mutual-Distillation |
title_full_unstemmed |
Knowledge Transfer via Dense Cross-Layer Mutual-Distillation |
title_sort |
knowledge transfer via dense cross-layer mutual-distillation |
publisher |
arXiv |
publishDate |
2020 |
url |
https://dx.doi.org/10.48550/arxiv.2008.07816 https://arxiv.org/abs/2008.07816 |
genre |
DML |
genre_facet |
DML |
op_rights |
arXiv.org perpetual, non-exclusive license http://arxiv.org/licenses/nonexclusive-distrib/1.0/ |
op_doi |
https://doi.org/10.48550/arxiv.2008.07816 |
_version_ |
1766397628905422848 |