Deep conditional adaptation networks and label correlation transfer for unsupervised domain adaptation. (February 2020)

• Record Type: Journal Article
• Authors: Chen, Yu; Yang, Chunling; Zhang, Yan; Li, Yuze
• Is Part Of: Pattern recognition. Volume 98(2020:Feb.)
• Publication Date: 2020-02
• Subjects: Conditional domain adaptation -- Deep learning -- Unsupervised learning -- Label transfer; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.107072
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Presenting a conditional adaptation networks for cross-domain image classification. Solving the categories mismatch and class prior bias problems by conditional adaptation. Proposing a label correlation transfer algorithm to preserve the domain information. Experiments are performed to show the usefulness of the proposed method. Abstract: Unsupervised domain adaptation aims to improve the performance of an unknown target domain by utilizing the knowledge learned from a related source domain. Given that the target label information is unavailable in the unsupervised situation, it is challenging to match the domain distributions and to transfer the source model to target applications. In this paper, a Deep Conditional Adaptation Networks (DCAN) is proposed to address the unsupervised domain adaptation problem. DCAN is implemented based on a deep neural network and attempts to learn domain invariant features based on the Wasserstein distance. A conditional adaptation strategy is presented to reduce the domain distribution discrepancy and to address category mismatch and class prior bias, which are usually ignored in marginal adaptation approaches. Furthermore, we propose a label correlation transfer algorithm to address the unsupervised issues, by generating more effective pseudo target labels based on the underlying cross-domain relationship. A set of comparative experiments were performed on standard domain adaptation benchmarks and the results demonstrate that the proposed DCAN outperforms previous adaptation methods.