Towards Semi-Supervised Causal Open Set Recognition 2022-01-7031

Most current deep learning methods assume the same class distributions for training and testing datasets. However, recognition of possible unknown class samples, i.e., classes not included in training that appear in testing, is common in the real world. This realistic problem is known as open-set recognition (OSR), where a classifier is trained to not only distinguish between known classes, but also to identify unknown classes as “unseen”. However, current state-of-the-art OSR methods rely heavily on large amounts of labeled training data, which are often not easily available in real applications. In this paper, we propose a novel semi-supervised causal open set recognition framework, which is motivated by the idea that generalized class and sample attributes learned through both labeled and unlabeled data will allow for the generation of more accurate counterfactuals, increasing the accuracy of unseen and seen recognition. Based on the proposed framework, a novel counterfactual contrastive loss is designed to increase the consistency of counterfactual generation across labeled and unlabeled data. Extensive experiments conducted across five datasets demonstrate that our method outperforms state-of-the-art methods. We show that bridging the gap between semi-supervised learning methods and causal-based generative models mark a significant advancement in OSR by utilizing both limited amounts of expensive, labeled data and a much larger, inexpensive collection of unlabeled data.