The most competitive noisy label learning methods rely on an unsupervised\nclassification of clean and noisy samples, where samples classified as noisy\nare re-labelled and "MixMatched" with the clean samples. These methods have two\nissues in large noise rate problems: 1) the noisy set is more likely to contain\nhard samples that are in-correctly re-labelled, and 2) the number of samples\nproduced by MixMatch tends to be reduced because it is constrained by the small\nclean set size. In this paper, we introduce the learning algorithm PropMix to\nhandle the issues above. PropMix filters out hard noisy samples, with the goal\nof increasing the likelihood of correctly re-labelling the easy noisy samples.\nAlso, PropMix places clean and re-labelled easy noisy samples in a training set\nthat is augmented with MixUp, removing the clean set size constraint and\nincluding a large proportion of correctly re-labelled easy noisy samples. We\nalso include self-supervised pre-training to improve robustness to high noisy\nlabel scenarios. Our experiments show that PropMix has state-of-the-art (SOTA)\nresults on CIFAR-10/-100(with symmetric, asymmetric and semantic label noise),\nRed Mini-ImageNet (from the Controlled Noisy Web Labels), Clothing1M and\nWebVision. In severe label noise bench-marks, our results are substantially\nbetter than other methods. The code is available\nathttps://github.com/filipe-research/PropMix.\n