Learning Likelihood Estimates for Open Set Domain Adaptation

Most existing domain adaptation frameworks were based on the strict assumption that different domains share the same label space, which is too idealized for real-world applications. In this paper, we examine a realistic case of open set domain adaptation with partially shared classes in two domains and consider the negative transfer brought by the unknown class from the target domain. Therefore, how to distinguish the unknown class from the known ones plays a key role. To achieve this goal, we propose a new approach called Open Distribution Control Adaptation Network(OpenDCAN). Inspired by the idea of Out-of-Distribution Detection, our method tries to explicitly model the distribution of the known source samples to differentiate it from unknown class in probabilistic distribution, which is different from existing methods. We use OpenBP, an open set domain adaptation framework, to provide a rough decision boundary between known and unknown classes. Then we further boost the performance by modifying the recently proposed Adaptive Feature Norm algorithm for the open set setting. If the probabilistic distribution of source samples is modeled, a likelihood estimate can easily detect unknown class without requiring any unknown samples in the source domain to get the concept of “unknown”. Experiments on representative benchmark datasets demonstrate the effectiveness of our approach.