Universal Domain Adaptation

Domain adaptation with strict assumptions on the label set relations is extensively explored both in previous chapters and in the literature. These assumptions simplify the adaptation problem and help researchers focus on the domain gap, yet limit the practical value of domain adaptation. In this chapter, we introduce Universal Domain Adaptation (UDA) which requires no prior knowledge about label sets in the source domain and the target domain. Given a pair of label sets (source and target), their union set can be divided into three parts: two label sets private to each domain and a common label set (the target label set is invisible to the model during training). An additional category gap across domains naturally arises under the UDA setup. Since the target label set is divided into two parts (target private labels and common labels), any model for UDA should make a choice for each input: (1) either classify it into the target private labels (unified as a single “unknown” class), (2) or recognize that its label belongs to the common label set and therefore predict its specific label. The UDA setting fits well in domain adaptation in the wild, where we have no knowledge about the target label set. Hereby we also propose Universal Adaptation Network (UAN) to solve the practical UDA problem. The main idea is to design a transferability criterion to identify labels common and private to each domain, thereby promoting the adaptation across common labels and identifying the “unknown” examples with low transferability.