Asymmetric alignment joint consistent regularization for multi-source domain adaptation
- PDF / 1,933,917 Bytes
- 24 Pages / 439.642 x 666.49 pts Page_size
- 83 Downloads / 239 Views
Asymmetric alignment joint consistent regularization for multi-source domain adaptation Junyuan Shang1 · Chang Niu1 · Zhiheng Zhou1 Xiangwei Li3
· Junchu Huang1 · Zhiwei Yang2 ·
Received: 23 June 2020 / Revised: 29 August 2020 / Accepted: 15 September 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Most existing methods of domain adaptation are proposed for single-source settings, where the source data come from a single domain. However, in practice, the available data generally come from multiple domains with different distributions. In this paper, we propose asymmetric alignment joint consistent regularization (AACR) for multi-source domain adaptation. As for asymmetric alignment, we propose to learn asymmetric projections, explicitly treating each domain differently to avoid the loss of shared information. Data from different domains are encoded into a common subspace by these asymmetric projections, where the encoded features are further aligned to promote knowledge transfer. Further, we propose consistent regularization to better learn shared information and filter out domainspecific information. We formulate the two parts into a unified framework, and derive its global optimal solution. Comprehensive experiments are conducted to evaluate AACR, and the results verify the effectiveness and robustness of our method. Keywords Subspace learning · Domain adaptation · Feature extraction · Image classification
1 Introduction As a basic task in the field of image processing and computer vision [4, 8–10, 25], image classification [5, 6, 12] has achieved significant advances in recent years. For classification tasks, labeled information takes a pivotal place in training effective model. Unfortunately, with the continuous emergence of new applications, we can only obtain limited or even no labeled data. Moreover, manually labeling data is expensive and time-consuming. One may Zhiheng Zhou
[email protected] 1
School of Electronic and Information Engineering, South China University of Technology, GuangZhou, China
2
China Information and Communication Research Institute, Guangzhou, People’s Republic of China
3
School of Computer & Artificial Intelligence, Lanzhou Institute of Technology, Lanzhou, People’s Republic of China
Multimedia Tools and Applications
expect to utilize labeled data in related source domains to facilitate learning of unlabeled target domains. However, due to the existence of domain shift [41], applying the model trained on the source domains directly to target domains may lead to poor performance. To solve this problem, domain adaptation [2, 36, 37] is proposed, aiming to promote knowledge transfer by reducing the domain discrepancies [35]. Although many works [20, 23, 48] have made great progress in domain adaptation, most of them concentrate on single-source settings where the source data come from one domain. Compared with single-source domain adaptation, multi-source domain adaptation has higher application value, as the available data generally c
Data Loading...