- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Unsupervised domain adaptation and generative model...
Open Collections
UBC Theses and Dissertations
UBC Theses and Dissertations
Unsupervised domain adaptation and generative model for intelligent engineering systems Wang, Jing
Abstract
Data-driven approaches have been widely applied to intelligent engineering systems. Their performance has been improved significantly by deep neural networks, which take advantage of large quantities of labeled data. However, the discriminative model that is trained using one dataset typically performs poorly when another different but related dataset is applied. In other words, it is difficult to design a discriminative model that can perform consistently well in different application scenarios without using any adaptation strategies. To tackle this problem, a research field called unsupervised domain adaptation, which aims at transferring knowledge from a label-rich source domain to an unlabeled target domain becomes active, has emerged. It is known that a discriminative model that is independent of the users or the environments, can be generalized by using unsupervised domain adaptation algorithms. However, there is a problem with all existing unsupervised domain adaptation methods. They cannot always learn a common representation space for the features from the two domains, making it difficult for the target domain to take advantage of the discriminative source features for its classification. To tackle this problem, specifically for engineering applications, two novel approaches, namely discriminative feature alignment and mutual variational inference, are proposed in the thesis. The proposed discriminative feature alignment can ensure that the features from the two domains can be properly constructed in a single distribution space, which is the space of a predefined Gaussian prior distribution where the target input samples can maximally take advantage of the discriminative source features for their classification tasks. The proposed mutual variational inference can indirectly transfer the knowledge learned from the source domain to the target domain via variational inference and mutual-information optimization. The developments are implemented on practical engineering problems and the performance is carefully evaluated to validate them.
Item Metadata
Title |
Unsupervised domain adaptation and generative model for intelligent engineering systems
|
Creator | |
Publisher |
University of British Columbia
|
Date Issued |
2020
|
Description |
Data-driven approaches have been widely applied to intelligent engineering systems. Their performance has been improved significantly by deep neural networks, which take advantage of large quantities of labeled data. However, the discriminative model that is trained using one dataset typically performs poorly when another different but related dataset is applied. In other words, it is difficult to design a discriminative model that can perform consistently well in different application scenarios without using any adaptation strategies. To tackle this problem, a research field called unsupervised domain adaptation, which aims at transferring knowledge from a label-rich source domain to an unlabeled target domain becomes active, has emerged. It is known that a discriminative model that is independent of the users or the environments, can be generalized by using unsupervised domain adaptation algorithms. However, there is a problem with all existing unsupervised domain adaptation methods. They cannot always learn a common representation space for the features from the two domains, making it difficult for the target domain to take advantage of the discriminative source features for its classification. To tackle this problem, specifically for engineering applications, two novel approaches, namely discriminative feature alignment and mutual variational inference, are proposed in the thesis. The proposed discriminative feature alignment can ensure that the features from the two domains can be properly constructed in a single distribution space, which is the space of a predefined Gaussian prior distribution where the target input samples can maximally take advantage of the discriminative source features for their classification tasks. The proposed mutual variational inference can indirectly transfer the knowledge learned from the source domain to the target domain via variational inference and mutual-information optimization. The developments are implemented on practical engineering problems and the performance is carefully evaluated to validate them.
|
Genre | |
Type | |
Language |
eng
|
Date Available |
2020-05-11
|
Provider |
Vancouver : University of British Columbia Library
|
Rights |
Attribution-NonCommercial-NoDerivatives 4.0 International
|
DOI |
10.14288/1.0390451
|
URI | |
Degree | |
Program | |
Affiliation | |
Degree Grantor |
University of British Columbia
|
Graduation Date |
2020-11
|
Campus | |
Scholarly Level |
Graduate
|
Rights URI | |
Aggregated Source Repository |
DSpace
|
Item Media
Item Citations and Data
Rights
Attribution-NonCommercial-NoDerivatives 4.0 International