Beyond Classical Domain Adaptation

Abstract

To develop learning methods across domains it is usual to make some simplifying assumptions. In standard domain adaptation the source consists of samples drawn from a single data distribution. The same holds also for the target, and the two domains share exactly the same label set. However, in the real world those conditions are often violated and several studies have formalized the sub-problems that arise when relaxing these assumptions. For example, when increasing the number of available domains (multi-source and multi-target DA), or reducing the access to manual annotation (zero-shot DA, predictive DA, online DA) or to source data (source-free and federated DA). The label set might also be different between source and target with more classes in the source (partial DA), in the target (open-set DA) or a more general a partial overlap combining these cases (universal DA). In this chapter we provide an overview of the techniques proposed to tackle all those challenging scenarios.