This paper proposed a novel information-theoretic loss function for learning with noisy labels. Besides regularizations borrowed from other areas, current directions in this area include loss correction, sample selection/reweighting, and label correction; different from existing loss-correction methods, the proposed loss is theoretically insensitive to label noise. This novelty is the reason for an acceptance recommendation. However, the authors didn't address well the reviewers' concerns in the rebuttal. I called the 4th reviewer to perform a quick review, and this reviewer is happy with the idea but again unhappy with the experiments done so far. Note that this acceptance is only conditional, so please carefully follow the reviews to revise the manuscript and prepare the final version.