Detecting Noisy Training Data with Loss Curves

This paper introduces a new method to discover mislabeled training samples and to mitigate their impact on the training process of deep networks. At the heart of our algorithm lies the Area Under the Loss (AUL) statistic, which can be easily computed for each sample in the training set. We show that the AUL can use training dynamics to differentiate between (clean) samples that benefit from generalization and (mislabeled) samples that need to be “memorized”. We demonstrate that the estimated AUL score conditioned on clean vs. noisy is approximately Gaussian distributed and can be well estimated with a simple Gaussian Mixture Model (GMM). The resulting GMM provides us with mixing coefficients that reveal the percentage of mislabeled samples in a data set as well as probability estimates that each individual training sample is mislabeled. We show that these probability estimates can be used to down-weight suspicious training samples and successfully alleviate the damaging impact of label noise. We demonstrate on the CIFAR10/100 datasets that our proposed approach is significantly more accurate and consistent across model architectures than all prior work.