Weighted progressive alignment for multi-source domain adaptation
Lists
Only you see this
Abstract
Multi-source domain adaptation (MSDA) dedicates to establishing knowledge transfer from multiple labeled source domains to an unlabeled target domain. Although data from multiple source domains can provide rich information, it also brings two problems. First, it is not easy to directly align multiple source domains and target domain, because there are complex interactions among multiple source domains with different distributions. Second, some of the source samples may contribute negatively to domain adaptation. Thus, how to select the appropriate source domain samples is worth exploring. To solve these problems, we propose a novel framework of weighted progressive alignment (WPA), in which we develop a two-stage alignment with two distinct domain classifiers, as well as a dedicated classifier to judge the importance of source domain samples. Our proposed method progressively achieves multi-source domain adaptation through domain-adversarial training and coarse-to-fine alignment. We evaluate our framework on four public benchmark datasets. The extensive experimental results demonstrate that the proposed method achieves great performance.
BibTeX
@article{wu2023weighted,
author = {Kunhong Wu and Liang Li and Yahong Han},
title = {{Weighted progressive alignment for multi-source domain adaptation}},
journal = {Multim. Syst.},
year = {2023},
doi = {10.1007/s00530-022-00987-7},
}Loading related content...