Model-Agnostic Label Quality Scoring to Detect Real-World Label Errors

We consider algorithms to find wrongly labeled data, which lurks in many real-world applications and hampers training/evaluation of ML models. We present the first empirical study of various scoring methods for this task on real datasets with naturally-occurring label errors (as opposed to synthetically introduced label errors). The label quality scores considered here can be utilized with arbitrary classification models. We examine five popular image recognition models (and ensembles thereof) to comprehensively characterize how well different scores detect label errors in practice.