Mutual Filter Teaching for Open-Set Semi-Supervised Learning

Open-set semi-supervised learning (OSSL) provides a practical solution by filtering out-of-distribution (OOD) samples from unlabeled data to guarantee the reliance on large unlabeled data in semi-supervised setting. However, existing OSSL methods mainly focus on identifying in-distribution (ID) samples and discarding OOD samples, while ignoring to make full use of samples that could not be exactly identified as ID or OOD samples. Those samples are more likely to be hard samples, which should be carefully explored to boost the performance in OSSL task. Hence, in this paper, we propose a novel framework, named Mutual Filter Teaching (MFT), where two networks are trained simultaneously to divide the unlabeled data into three parts: ID samples, OOD samples and hard samples. The samples are regarded as ID or OOD samples only if two networks give consistent decisions according to Mahalanobis distance between the unlabeled samples and their closest class prototypes. For those samples with inconsistent decisions, we treat them as hard samples and design an efficient mutual teaching scheme where the samples detected by only one network as positive samples are fed to its peer network for training. Furthermore, we propose to employ the prediction variance of two networks to dynamically rectify the learning from hard samples. Experiments on multiple benchmark datasets demonstrate that our approach achieves the state-of-the-art performance.