Identifying Mislabeled Data using the Area Under the Margin Ranking
This addresses data quality issues for machine learning practitioners by providing a method to clean training datasets, though it appears incremental as it builds on prior work in mislabel detection.
The paper tackles the problem of mislabeled data in training sets by introducing the Area Under the Margin (AUM) statistic to identify such samples, resulting in improved test error rates, such as a 1.6% absolute improvement on WebVision50 and a 1.2% drop on CIFAR100 after removing portions of the data.
Not all data in a typical training set help with generalization; some samples can be overly ambiguous or outrightly mislabeled. This paper introduces a new method to identify such samples and mitigate their impact when training neural networks. At the heart of our algorithm is the Area Under the Margin (AUM) statistic, which exploits differences in the training dynamics of clean and mislabeled samples. A simple procedure - adding an extra class populated with purposefully mislabeled threshold samples - learns a AUM upper bound that isolates mislabeled data. This approach consistently improves upon prior work on synthetic and real-world datasets. On the WebVision50 classification task our method removes 17% of training data, yielding a 1.6% (absolute) improvement in test error. On CIFAR100 removing 13% of the data leads to a 1.2% drop in error.