Deep k-NN for Noisy Labels

The dataset used in the paper is not explicitly described, but it is mentioned that the authors used a preliminary model to compute an intermediate representation and then applied k-NN filtering to remove mislabeled training data.