Reviews: Using Trusted Data to Train Deep Networks on Labels Corrupted by Severe Noise

Summary A method for learning a classifier robust to label noise is proposed. In contrast with most previous work, the authors take a simplifying but realistic assumption: there is of a small subset of the training set that has been sanitized (where labels are not noisy). Leveraging prior work, a noise transition matrix is first estimated and then used to correct the loss function for classification. Very extensive experiments support the validity of the method. Detailed comments This is a good contribution, although rather incremental with respect to Patrini et al. '17.