Learning transferable and discriminative features for unsupervised domain adaptation

Although achieving remarkable progress, it is very difficult to induce a supervised classifier without any labeled data. Unsupervised domain adaptation able overcome this challenge by transferring knowledge from source an unlabeled target domain. Transferability and discriminability are two key criteria for characterizing the superiority of feature representations enable successful adaptation. In paper, novel method called learning TransFerable Discriminative Features unsupervised (TFDF) proposed optimize these objectives simultaneously. On one hand, distribution alignment performed reduce discrepancy learn more transferable representations. Instead adopting Maximum Mean Discrepancy (MMD) which only captures first-order statistical information measure discrepancy, we adopt recently statistic Covariance (MMCD), can not capture but also second-order in reproducing kernel Hilbert space (RKHS). other propose explore both local discriminative via manifold regularization global minimizing class confusion objective features, respectively. We integrate into Structural Risk Minimization (RSM) framework domain-invariant classifier. Comprehensive experiments conducted on five real-world datasets results verify effectiveness method.