Barely-Supervised Learning: Semi-supervised Learning with Very Few Labeled Images

This paper tackles the problem of semi-supervised learning when set labeled samples is limited to a small number images per class, typically less than 10, that we refer as barely-supervised learning. We analyze in depth behavior state-of-the-art method, FixMatch, which relies on weakly-augmented version an image obtain supervision signal for more strongly-augmented version. show it frequently fails scenarios, due lack training no pseudo-label can be predicted with high confidence. propose method leverage self-supervised methods provides absence confident pseudo-labels. then two refine selection process lead further improvements.The first one per-sample history model predictions, akin voting scheme. The second iteratively up-dates class-dependent confidence thresholds better explore classes are under-represented Our experiments our approach performs significantly STL-10 regime,e.g. 4 or 8 class.