Domain generalization based on domain-specific adversarial learning

Deep learning models often suffer from degraded performance when the distributions of the training and testing data differ (i.e., domain shift). Domain generalization (DG) techniques can help improve the generalization performance for unseen target domains by using multiple source domains. The recently developed domain generalization methods focus on extracting domain-invariant features from all source domains. However, some task-relevant discriminative information can be removed during this process. In addition, the various source domains are treated equally ignoring the negative impacts of distant source domains. Both problems can lead to unsatisfactory performance. This paper proposed a domain-specific adversarial neural network (DSANN) based on adversarial learning to learn effective feature representations and reduce the influence of distantsource domains. The DSANN introduces a reference distribution that is adaptively generated during training. Additionally, domain-invariant features are extracted through a domain-specific adversarial learning process , in which each source domain distribution is aligned only with the reference distribution instead of all the other source domains. Moreover, the DSANN also aligns the outputs of multiple classifiers and adopts the weighted average of the predictions; thus, the employed label classifiers can become more robust to unknown domain shifts. Experiments conducted on popular benchmark datasets demonstrate that our proposed method can achieve remarkable generalization performance and has better classification accuracy than the existing DG algorithms.