Discriminative Adversarial Domain Adaptation Based on Circle Loss

Transfer learning can solve some problems that the lack of data and labeled data is unavailable in deep learning. The domain adaptation method in transfer learning can solve the problem of domain-shift, which is the main direction of current research. The generative adversarial network is popular in the generative model now, So the combination of them will obtain more respectable performance. While previous approaches only use it as a data augmentation tool. Even though the overall distribution is aligned, some samples will still be misclassified, and the corresponding classifier is trained on the source samples and cannot perfectly generalize to the target. We propose a novel generative discriminative domain adaptation method based on the reverse gradient by combining the generative adversarial method to reverse gradient. The method uses unlabeled target domain data to make the target and source domains distributions get closer in the feature space. At the same time, the auxiliary class's generation adversarial network is used to generate consistent-class representation, and the corresponding class's distribution is aligned across domains. Besides, we use circle loss as the classification function. Circle loss proposed a unified formula for deep feature learning features. It is useful for our classification task in our method. Our proposed method achieves better performance than others in most experimental settings and has shown to work well across different datasets such as digits and Office.