Dual Adversarial Training for Unsupervised Domain Adaptation

Deep neural networks obtain remarkable achievements in diverse real-world applications. However, their success relies on the availability of large amounts of labeled data. A trained model may fail to generalize well on a domain whose distribution differs from the training data distribution. Collecting abundant labeled data for all domains of interest are expensive and time-consuming, sometimes even impossible. Domain adaptation sets out to address this problem, aiming to leverage labeled data in the source domain to learn a good predictive model for the target domain whose labels are scarce or unavailable. A mainstream approach is adversarial domain adaptation, which learns domain invariant-features by performing alignment across different distributions. Most domain adaptation methods focus on reducing the divergence between two domains to make the improvement. A prerequisite of domain adaptation is the adaptability, which is measured by the expected error of the ideal joint hypothesis on the source and target domains, should be kept at a small value in the process of domain alignment. However, adversarial learning may degrade the adaptability, since it distorts the original distributions by suppressing the domain-specific information. In this paper, we propose an approach, which focuses on strengthening the model's adaptability, for domain adaptation. Our proposed dual adversarial training (DAT) method introduces class-invariant features to enhance the discriminability of the latent space without sacrificing the transferability. The class-invariant features, extracted from the source domain, can play a positive role in the classification on the target domain. We demonstrate the effectiveness of our method by yielding state-of-the-art results on several benchmarks.