Unsupervised Domain Adaptation with Joint Domain-Adversarial Reconstruction Networks

摘要

Unsupervised Domain Adaptation (UDA) attempts to transfer knowledge from a labeled source domain to an unlabeled target domain. Recently, domain-adversarial learning has become an increasingly popular method to tackle this task, which bridges source domain and target domain by adversarially learning domain-invariant representations that cannot be discriminated by a domain discriminator. In spite of the great success achieved by domain-adversarial learning, most of existing methods still suffer two major limitations: (1) due to focusing only on learning domain-invariant representations, they ignore the individual characteristics of each domain and fail to extract domain-specific information that is beneficial for final classification; (2) by focusing only on performing domain-level distribution alignment to learn domain-invariant representations, they fail to achieve the invariance of representations at a class level, which may lead to incorrect distribution alignment. To address the above issues, we propose in this paper a novel model called Joint Domain-Adversarial Reconstruction Network (JDARN), which integrates domain-adversarial learning with data reconstruction to learn both domain-invariant and domain-specific representations. Meanwhile, we propose to employ two novel discriminators called joint domain-class discriminators to achieve the joint alignment and adopt a novel joint adversarial loss to train them. With both domain and class information of two domains, the two discriminators can be used to promote domain-invariant representation learning towards the class level, not only the domain level. Extensive experimental results reveal that the proposed JDARN exceeds the state-of-the-art performance on two standard UDA datasets.