| Author: | Mao, Sitong |
| Title: | Leveraging deep representations in layered neural networks for domain adaptation |
| Advisors: | Chung, Fu-lai Korris (COMP) |
| Degree: | Ph.D. |
| Year: | 2021 |
| Subject: | Adaptive computing systems Deep learning (Machine learning) Transfer learning (Machine learning) Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Computing |
| Pages: | 15, 133 pages : color illustrations |
| Language: | English |
| Abstract: | Domain adaptation aims to build a model for the target data with the help of the source data which comes from a related but different domain. This results in a difference of distributions between source data and target data which is referred to as domain shift. Consequently, the model pre-trained on the source domain can not perform well when being applied to the target domain directly. Hence, it is important to develop domain adaptation methods to enhance the performance. Our works explore domain adaptation methods under two different settings: the closed set domain adaptation and the open set domain adaptation. In the closed set domain adaptation, the target domain and the source domain contain the same set of categories while for the open set domain adaptation, the target domain contains target-specific categories that do not exist in the source domain. The first two works of this thesis focus on the closed set domain adaptation while the third work concerns with the open set domain adaptation. In the fourth work, a new domain adaptation setting is proposed. The domain adaptation problem has been recently addressed by employing adversarial learning and distinctive adaptation performance has been reported. In our second work, a deep adversarial domain adaptation model based on a multi-layer joint kernelized distance metric is proposed. By utilizing the abstract features extracted from deep networks, the multi-layer joint kernelized distance (MKD) between the jth target data predicted as the mth category and all the source data of the m'th category is computed. Base on MKD, a class-balanced selection strategy is utilized in each category to select target data that are most likely to be classified correctly and treat them as labeled data using their pseudo labels. Then, an adversarial architecture is used to draw the newly generated labeled training data and the remaining target data close to each other. In this way, the target data themselves provide valuable information to enhance the domain adaptation. An analysis of the proposed method is also given and the experimental results demonstrate that the proposed method can achieve a better performance than a number of state-of-the-art methods. Though adversarial training has achieved significant performance, it is still influenced by the mislabeled target data. In classification tasks, some target data with ground truth label "k" may be predicted as an incorrect category "k" because of domain shift. To address this problem, we propose a method to utilize information provided by the target domain to help re-examine the predicted results. Consequently, an auxiliary network is exploited to accumulate the target data predicted results of each iteration and thus the term "target-ensemble" is used to characterize the proposed method. By training on the target data alone, the auxiliary network can alleviate the influence of the source data and reflect more information of the target domain. By leveraging the information provided by the auxiliary network during the adversarial training process, the domain adaptation performance can be enhanced. Experimental results show that the proposed method can outperform several state-of-the-art domain adaptation algorithms. Our third work explores methods for open set domain adaptation which is a more application scenario compared with the typical closed set domain adaptation. The main difficulty in open set domain adaptation is that we need to distinguish which target data belongs to the unknown classes. We propose an "against adversarial learning" method that can distinguish unknown target data naturally without setting any additional hyper-parameter and the target data predicted to the known classes can be classified at the same time. A weighted adversarial architecture and a dynamic classifier classify all target data as unknown at first are utilized. Experimental results show that the proposed method can make significant progress in performance compared with several state-of-the-art methods. In the settings of conventional domain adaptation, categories of the source dataset are from the same domain (or domains for multi-source domain adaptation), which is not always true in reality. In our fourth work, we propose a new setting named "Inter-Category Domain Shift" (ICDS). Under the settings of ICDS, different categories of the source dataset are collected from several domain. Under such situation, domain adaptation performance will be further influenced because of the distribution discrepancy inside the source data. We analyze the influence from both empirical and theoretical perspective. A feature element-wise weighting (FEW) method that can reduce distribution discrepancy between different categories is also proposed. Various experiments are conducted to show the significance of ICDS and the effectiveness of FEW. The proposed ICDS and FEW can be easily extended to the most special case where each source category comes from a different domain. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/11789