Cross-view classification by joint adversarial learning and class-specificity distribution

Abstract Despite the promising preliminary results, none of existing deep learning based cross-view classification methods simultaneously takes into account both view consistency learning and class-specificity distribution of the extracted features, resulting in unstable classification performance. Moreover, most existing cross-view classification methods are sensitive to scale due to the scale issue of view representations, resulting in unstable view-consistent representations. In this paper, we propose a new deep adversarial network for cross-view classification that attempts to learn robust view-consistent representations by combing the thought of adversarial learning and metric learning in Fisher criterion. Meanwhile, a class-specificity distribution term, which is measured by l12-norm, is employed to make the view-consistent representations with the same label to further have a common distribution in dimension space while view-representations with different labels have different distribution in the intrinsic dimension space. We formulate the aforementioned two concerns into a unified optimization framework. Extensive experiments on several real-world datasets indicate the effectiveness of our method over the other state-of-the-arts.