Review for NeurIPS paper: Dual T: Reducing Estimation Error for Transition Matrix in Label-noise Learning

Weaknesses: 1) The justification and explanation of equation 3 (which is a central point of the paper) is not clear. Here's how I interpret the approach proposed by the authors based on Algorithm 1. The elements of the Dual T-estimator transition matrix are computed as follows: \hat{T}_{ij} \sum_l \hat{P}(\bar{Y} j Y' l) \hat{P}(Y' l Y_i) The second element in the sum is obtained using equation 1, which is the same equation used to compute the T-estimator of the transition matrix. The first element in the sum is estimated using equation 4 which determines the number of examples belonging to the noisy class j which were incorrectly labeled as belonging to the noisy class l divided by the number of examples labeled as belonging to the noisy class l. This ratio is basically used to revise / correct the T-estimator of the transition matrix (and hence mitigate the effect of overfitting the noise in the training set).