Robust Semi-Supervised Learning with Out of Distribution Data

Recent Semi-supervised learning (SSL) works show significant improvement in SSL algorithms' performance using better-unlabeled data representations. However, recent work [Oliver et al., 2018] shows that the SSL algorithm's performance could degrade when the unlabeled set has out-of-distribution examples (OODs). In this work, we first study the critical causes of OOD's negative impact on SSL algorithms. We found that (1) the OOD's effect on the SSL algorithm's performance increases as its distance to the decision boundary decreases, and (2) Batch Normalization (BN), a popular module, could degrade the performance instead of improving the performance when the unlabeled set contains OODs. To address the above causes, we proposed a novel unified-robust SSL approach that can be easily extended to many existing SSL algorithms, and improve their robustness against OODs. In particular, we propose a simple modification of batch normalization, called weighted batch normalization, that improves BN's robustness against OODs. We also developed two efficient hyper-parameter optimization algorithms that have different tradeoffs in computational efficiency and accuracy. Extensive experiments on synthetic and real-world datasets prove that our proposed approaches significantly improves the robustness of four representative SSL algorithms against OODs compared with four state-of-the-art robust SSL approaches.