Deep Discriminative Domain Adaptation

Domain adaptation studies the problem of how to transfer knowledge across different domains where the source domain with rich labeled source samples and target domain with less or even no labeled target samples are drawn from different probability distribution. A prevailing strategy is to generate transferable features cross-domain, which induces deficient knowledge transfer in learning process, especially close to the input end. Moreover, when target samples distributed far their corresponding class centers, or near the edge of the clusters, only using learned source samples features to predict target samples class which may easily brings about misclassification. Targeting to deal with these issues, we propose Deep Discriminative Domain Adaptation (DDDA) method, which jointly minimizes the supervised classification loss of annotated source examples, the unsupervised center alignment and correlation alignment losses measured on both convolutional layers and fully connected layers with help of attention mechanism. The multi-layer transfer mechanism complementary strengthens each individual transfer component, and markedly improves the generalization ability of transfer models. A series of experiments conducted on several standard datasets validate that the proposed method consistently outperforms contemporary adaptation approaches.