Domain adaptation based on source category prototypes

Unsupervised domain adaptation (UDA), which can transfer knowledge from labeled source domain to unlabeled target domain, needs to access a large number of labeled source data in the process of generalization. However, the data of two domains may not be accessed at the same time due to data privacy protection. To solve this problem, source-data free domain adaptation (SFDA) began to receive attention. However, too little source information will lead to some performance gaps. To balance the issues between UDA and SFDA, a new setting called Prototype-based domain adaptation (Prototype-DA) is proposed, which further improves the practicability of UDA by using source category prototype instead of source data. At the same time, it can also ensure the privacy of source data like SFDA. Specifically, our training process can be divided into two steps. First, the source data is used to pre-train a source model, and the source category prototypes are obtained after the training of source model. Then, to generalize the source model to the target domain, category maximum mean discrepancy (Category-MMD) is defined so that the target data can be aligned with the source category prototypes. In this way, source category prototypes will transfer knowledge to the target domain together with the source model. Through source category prototypes, Prototype-DA can not only achieve the comparable results than the method using source data, but also protect the privacy of source data to some extent. Furthermore, the target category prototypes are constructed and the consistency between the labels of target category prototypes and the classification results is required. This prototype-label consistency regularization, proposed by us for the first time, helps to extract discriminative features in the target domain. Compared with the previous UDA methods and SFDA methods, extensive experiments on multiple public domain adaptation datasets show that Prototype-DA achieves the state-of-the-art results. At the same time, the traditional UDA theory is expanded to our method setting and makes a theoretical analysis to ensure the effectiveness of our method.