Adversarial Learning Approach for Open Set Domain Adaptation

Many methods have been proposed for adapting a model trained in a label-rich domain (source) to a label-scarce domain (target). These methods have the assumption that these domains completely have the same the categories. However, if examples in target domain are not given label, we cannot make sure that the domains share the category. A target domain may include examples of categories that the source domain does not have (open set domain adaptation), or some categories can be absent in the target domain (partial domain adaptation). Methods that perform well in this situation are very useful. In this chapter, we briefly summarize non-closed domain adaptation settings in the related work and introduce a method for open set domain adaptation. We define the shared class as the known class and the unshared class as the unknown class. Most existing distribution matching based methods do not work well in the open set situation because unknown target samples should not be matched with the source. In this chapter, we introduce a method which utilizes adversarial training (Saito et al. (Open set domain adaptation by backpropagation. In: Proceedings of the European Conference on Computer Vision (ECCV), 2018)). A classifier is trained to make a boundary between the source and the target samples whereas a generator is trained to make target samples far from the boundary. The key idea of the method is to assign two options to the feature generator: aligning them with source known samples or rejecting them as unknown samples. This approach allows to extract features that separate unknown target samples from known target samples.