Few-shot Out-of-Distribution Detection

Out-of-distribution (OOD) detection is an important problem in real-world settings, and has inspired a wide range of methods, from simple ones based on the predicted probability of a classifier to more complicated ones based on likelihood ratios under deep generative models. We consider a variant of the OOD detection task appropriate to settings such as few-shot learning, in which classification involves a restricted set of novel classes. We establish baselines on the few-shot OOD detection tasks by adapting state-of-the-art OOD methods from the standard classification setting to the few-shot setting. Interestingly, strong baselines designed for the large data setting perform well in the few-shot setting after simple adaptation. Then we present a method for FS-OOD detection that specifically utilizes the structure of the few-shot problem, and show that it outperforms the previous methods Furthermore, we demonstrate that improvements in few-shot OOD detection can benefit downstream tasks, such as active learning and semi-supervised learning.