How AI describes DivideMix

DivideMix is a PyTorch implementation of the ICLR 2020 paper DivideMix: Learning with Noisy Labels as Semi-supervised Learning. It uses a per-sample loss mixture model to dynamically separate clean-label data from unlabeled data and then trains with semi-supervised learning to improve robustness to noisy labels. The repository provides code and experiments for datasets such as CIFAR, Clothing1M, and WebVision, authored by Junnan Li, Richard Socher, and Steven C.H. Hoi (Salesforce Research).