Abstract:
To collect large scale annotated data, it is inevitable to introduce label noise, i.e., incorrect class labels. A major challenge is to develop robust deep learning models that achieve high test performance despite training set label noise. We introduce a novel approach that directly cleans labels in order to train a high quality model. Our method leverages statistical principles to correct data labels and has a theoretical guarantee of the correctness. In particular, we use a likelihood ratio test to flip the labels of training data. We prove that the corrected labels are consistent with the true Bayesian optimal classifier with high probability. We incorporate our label correction algorithm into the training of deep neural networks and train models that achieve superior testing performance on multiple public datasets.
