Improving Barely Supervised Learning by Discriminating Unlabeled Samples with Super-Class

This paper addresses the problem of model collapse under very few labels (one label per class) and proposes superclass mining for potential discriminative information to aid model training. Download paper here