Unsupervised Domain Adaptation via bidirectional generation and middle domains alignment. (September 2022)

• Record Type: Journal Article
• Authors: Tian, Qing; Zhu, Yanan; Sun, Heyang; Yang, Hong; Xu, Heng
• Is Part Of: Computers & electrical engineering. Volume 102(2022)
• Publication Date: 2022-09
• Subjects: Unsupervised domain adaptation (UDA) -- Bidirectional generation -- Fake source/target domain -- Cross-domain alignment; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2022.108229
• Languages: English
• ISSNs: 0045-7906

Abstract

Abstract: The research involved Unsupervised Domain Adaptation (UDA) has huge emerged. However, almost of them assume that the domain distributions could be transformed into a shared representation space, which does not always hold, making domain-invariant representations challenging to generate directly on the original domains. We propose a novel UDA algorithm introducing a bidirectional generation and middle domains alignment (BGMA) that uses the idea of transmitting middle domains to reduce a cross-domain gap. To explore the middle domains (i.e., fake source/target domain), we embed supervised knowledge in the constructed task network to train a bidirectional generator. Besides, we replace the source and target domains with the generated middle domains, and we reassign the relative influence of each sample to the clusters' centers. Then we specifically design an algorithm to optimize our method with generalization bound analysis and demonstrate its effectiveness through a comprehensive evaluation. Graphical abstract: Highlights: A novel UDA method via bidirectional generation and middle domains alignment (BGMA) is proposed, which aligns discrepancies between the source and target domains through the constructed middle domains as the surrogate. Considering the classifiers' prediction of fake target domain is the biased estimation. We reasonably reassign the relative effect of each sample to the centers of the clusters. Our method jointly performs transmitting middles, supervised guidance of cross-domain and decision classifier more effectively to reduce the domain gap. To express the process of BGMA clearly, an optimization algorithm and the generalization error bound for BGMA are constructed. The effectiveness and superior performance of the BGMA are verified through our extensive ablation studies and evaluations on Office-31, Office-Home and ImageCLEF-DA datasets.