Learning smooth representations with generalized softmax for unsupervised domain adaptation

Domain adaptation aims at training accurate classifiers in a target domain by utilizing the data in a different but related source domain, which has made great progress in many aspects. Most of existing methods try to match the moments of data by mapping them to feature space, either first or second moment, so as to match the distribution. Despite their appeal, such models often fail to guarantee the obtained features can be well-classified. In this paper, we propose generalized softmax and smooth regularization to extract features and adapt classifiers simultaneously. Considering label matrix as special features, generalized softmax has more tolerance to the diversity of samples belonging to the same class. Smoothness regularization guarantees stronger robustness between target features and decision boundary. Finally, we evaluate our method on several standard benchmark datasets. Empirical evidence shows that the proposed method is comparable or superior to existing methods, and the same results based on two classification schemes indicate that the smoothness regularization is effective.