Synthetic Source Universal Domain Adaptation through Contrastive Learning

Universal domain adaptation (UDA) is a crucial research topic for efficient deep learning model training using data from various imaging sensors. However, its development is affected by unlabeled target data. Moreover, the nonexistence of prior knowledge of the source and target domain makes it more...

Descripción completa

Guardado en:

Detalles Bibliográficos
Autor principal:	Jungchan Cho
Formato:	article
Lenguaje:	EN
Publicado:	MDPI AG 2021
Materias:	universal domain adaptation contrastive learning classification deep learning Chemical technology TP1-1185
Acceso en línea:	https://doaj.org/article/7a576453521840dfb764b12bf8684110
Etiquetas:	Agregar Etiqueta Sin Etiquetas, Sea el primero en etiquetar este registro!

id	oai:doaj.org-article:7a576453521840dfb764b12bf8684110
record_format	dspace
spelling	oai:doaj.org-article:7a576453521840dfb764b12bf86841102021-11-25T18:57:18ZSynthetic Source Universal Domain Adaptation through Contrastive Learning10.3390/s212275391424-8220https://doaj.org/article/7a576453521840dfb764b12bf86841102021-11-01T00:00:00Zhttps://www.mdpi.com/1424-8220/21/22/7539https://doaj.org/toc/1424-8220Universal domain adaptation (UDA) is a crucial research topic for efficient deep learning model training using data from various imaging sensors. However, its development is affected by unlabeled target data. Moreover, the nonexistence of prior knowledge of the source and target domain makes it more challenging for UDA to train models. I hypothesize that the degradation of trained models in the target domain is caused by the lack of direct training loss to improve the discriminative power of the target domain data. As a result, the target data adapted to the source representations is biased toward the source domain. I found that the degradation was more pronounced when I used synthetic data for the source domain and real data for the target domain. In this paper, I propose a UDA method with target domain contrastive learning. The proposed method enables models to leverage synthetic data for the source domain and train the discriminativeness of target features in an unsupervised manner. In addition, the target domain feature extraction network is shared with the source domain classification task, preventing unnecessary computational growth. Extensive experimental results on VisDa-2017 and MNIST to SVHN demonstrated that the proposed method significantly outperforms the baseline by 2.7% and 5.1%, respectively.Jungchan ChoMDPI AGarticleuniversal domain adaptationcontrastive learningclassificationdeep learningChemical technologyTP1-1185ENSensors, Vol 21, Iss 7539, p 7539 (2021)
institution	DOAJ
collection	DOAJ
language	EN
topic	universal domain adaptation contrastive learning classification deep learning Chemical technology TP1-1185
spellingShingle	universal domain adaptation contrastive learning classification deep learning Chemical technology TP1-1185 Jungchan Cho Synthetic Source Universal Domain Adaptation through Contrastive Learning
description	Universal domain adaptation (UDA) is a crucial research topic for efficient deep learning model training using data from various imaging sensors. However, its development is affected by unlabeled target data. Moreover, the nonexistence of prior knowledge of the source and target domain makes it more challenging for UDA to train models. I hypothesize that the degradation of trained models in the target domain is caused by the lack of direct training loss to improve the discriminative power of the target domain data. As a result, the target data adapted to the source representations is biased toward the source domain. I found that the degradation was more pronounced when I used synthetic data for the source domain and real data for the target domain. In this paper, I propose a UDA method with target domain contrastive learning. The proposed method enables models to leverage synthetic data for the source domain and train the discriminativeness of target features in an unsupervised manner. In addition, the target domain feature extraction network is shared with the source domain classification task, preventing unnecessary computational growth. Extensive experimental results on VisDa-2017 and MNIST to SVHN demonstrated that the proposed method significantly outperforms the baseline by 2.7% and 5.1%, respectively.
format	article
author	Jungchan Cho
author_facet	Jungchan Cho
author_sort	Jungchan Cho
title	Synthetic Source Universal Domain Adaptation through Contrastive Learning
title_short	Synthetic Source Universal Domain Adaptation through Contrastive Learning
title_full	Synthetic Source Universal Domain Adaptation through Contrastive Learning
title_fullStr	Synthetic Source Universal Domain Adaptation through Contrastive Learning
title_full_unstemmed	Synthetic Source Universal Domain Adaptation through Contrastive Learning
title_sort	synthetic source universal domain adaptation through contrastive learning
publisher	MDPI AG
publishDate	2021
url	https://doaj.org/article/7a576453521840dfb764b12bf8684110
work_keys_str_mv	AT jungchancho syntheticsourceuniversaldomainadaptationthroughcontrastivelearning
_version_	1718410471100383232

Synthetic Source Universal Domain Adaptation through Contrastive Learning

Ejemplares similares