Open-set learning under covariate shift

Abstract

Open-set learning deals with the testing distribution where there exist samples from the classes that are unseen during training. They aim to classify the seen classes and recognize the unseen classes. Previous studies typically assume that the marginal distribution of the seen classes is fixed across the training and testing distributions. In many real-world applications, however, there may exist covariate shift between them, i.e., the marginal distribution of seen classes may shift. We call this kind of problem as open-set learning under covariate shift, aim to robustly classify the seen classes under covariate shift and be aware of the unseen classes.We present a new open-set learning framework with covariate generalization based on supervised contrastive learning, called SC–OSG, inspired by the latent connection between contrastive learning and representation invariance. Specifically, we theoretically justify supervised contrastive learning that could promote the conditional invariance of representations, a critical condition for covariate generalization. SC–OSG generates multi-source samples to promote the representation invariance and improve the covariate generalization. Based on this, we propose a detection score that is specific to the proposed training scheme. We evaluate the effectiveness of our method on several real-world datasets, on all of which we achieve competitive results with state-of-the-art methods.