They are Not Completely Useless: Towards Recycling Transferable Unlabeled Data for Class-Mismatched Semi-Supervised Learning

Semi-Supervised Learning (SSL) with mismatched classes deals the problem that classes-of-interests in limited labeled data are only a subset of massive unlabeled data. As result, classical SSL methods would be misled by which possessed To solve this problem, some recent divide to useful in-distribution (ID) and harmful out-of-distribution (OOD) data, among latter should particularly weakened. potential value contained OOD is largely overlooked. remedy defect, paper proposes “Transferable Recycling” (TOOR) method properly utilizes ID as well “recyclable” enrich information for conducting class-mismatched SSL. Specifically, TOOR treats have close relationship recyclable, employs adversarial domain adaptation project them space In other words, recyclability an datum evaluated its transferability, recyclable transferred so they compatible distribution known classes-of-interests. Consequently, our extracts more from than existing methods, it achieves improved performance demonstrated experiments on typical benchmark datasets.