Abstract: The maximum mean discrepancy (MMD), as a representative distribution metric between source domain and target domain, has been widely applied in unsupervised domain adaptation (UDA), where both domains follow different distributions, and the labels from source domain are merely available. However, MMD and its class-wise variants possibly ignore the intra-class compactness and inter-class separability, thus reducing discriminability of feature representation. This paper proposes a focally discriminative loss for unsupervised domain adaptation. This method endeavors to improve the discriminative ability of MMD from two aspects: (1) the weights are re-designed for MMD in order to align the distribution of relatively hard classes across domains; (2) a focally contrastive loss is explored to tradeoff the positive sample pairs and negative ones for better discrimination. The integration of both losses can not only make the intra-class features close, but also push away the inter-class features far from each other. Moreover, the improved loss is simple yet effective, and it can be extended to the network structure of the attention mechanism. Experiments on several domain adaptation datasets verify the effectiveness of the proposed method.

Key words: unsupervised domain adaptation, weighted maximum mean discrepancy, focally contrastive loss, attention mechanism