Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks | MIT logo
A team led by researchers at MIT CSAIL looked at 10 major datasets that have been cited over 100,000 times (including ImageNet and Amazon’s reviews dataset) and found a 3.4% average error rate across all datasets, including 6% for ImageNet. The team created a demo showcase the different datasets errors covering mislabeled images, mislabeled text sentiment and mislabeled audio of YouTube videos. Paper abstract: We identify label errors in the test sets of 10 of the most commonly-used computer vision, natural language, and audio datasets, and subsequently study the potential for these label errors to affect benchmark results. Errors in test sets are numerous and widespread: we estimate an average of 3.4% errors across the 10 datasets, where for example 2916 label errors comprise 6% of the ImageNet validation set. Putative label errors are identified using confident learning algorithms and then human-validated via crowdsourcing (54% of the algorithmically-flagged candidates are indeed erroneously labeled). Traditionally, machine learning practitioners choose which model to deploy based on test accuracy - our findings advise caution here, proposing that judging models over correctly labeled test sets may be more useful, especially for noisy real-world datasets. Surprisingly, we find that lower capacity models may be practically more useful than higher capacity models in real-world datasets with high proportions of erroneously labeled data. For example, on ImageNet with corrected labels: ResNet-18 outperforms ResNet50 if the prevalence of originally mislabeled test examples increases by just 6%. On CIFAR-10 with corrected labels: VGG-11 outperforms VGG-19 if the prevalence of originally mislabeled test examples increases by just 5%.