Reviews: Decoupling "when to update" from "how to update"

Summary The paper proposes a meta algorithm for training any binary classifier in a manner that is robust to label noise. A model trained with noisy labels will overfit them trained for too long. Instead, one can train two models at the same time, initialized at random, and update by disagreement: the updates are performed only when the two models' prediction differ, a sign that they are still learning from the genuine signal in the data (not the noise); and instead, defensively, if the models agree on their predictions and the respective ground truth label is different, they should not perform an update, because this is a sign of potential label noise. A key element is the random initialization of the models, since the assumption is that the two should not give the same prediction unless they are close to converge; this fits well with deep neural networks, the target of this work. The paper provides a proof of convergence in the case of linear models (updated with perceptron algorithm and in the realizable case) and a proof that the optimal model cannot be reach in general, unless we resort to restrictive distributional assumptions (this is nice since it also shows a theoretical limitation of the meta-algorithm).