Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks
curtis-northcutt, Anish Athalye and Jonas Mueller, NeurIPS 2021 (Datasets and Benchmarks track), arXiv 2103.14749. The measurement half of the data layer, where data-cascades is the practice half.
What they measured
Ten of the most-used test sets in computer vision, NLP and audio — ImageNet, CIFAR-10, MNIST, QuickDraw, Amazon Reviews, IMDB and others. Candidate label errors were flagged algorithmically with confident learning, then put to crowdworkers for human confirmation.
- At least 3.3% of labels are wrong on average across the ten test sets.
- At least 6% of the ImageNet validation set.
- About half the flagged candidates survived human review — 51% were confirmed as genuinely mislabeled, which is also the honest reading of how noisy the detector itself is.
These are test sets. The sets the field uses to decide which model is better.
Correcting the labels reverses model rankings
Correct the labels and the leaderboard can invert. On corrected ImageNet, ResNet-18 beats ResNet-50 once the prevalence of originally mislabeled test examples rises by just 6 percentage points; on corrected CIFAR-10, VGG-11 beats VGG-19 at just 5.
So a decision the field makes constantly — take the bigger model, it scored higher — can be an artifact of the errors in the ruler. The larger model was partly being rewarded for reproducing the mistakes in the test set, and the smaller one penalized for not reproducing them.
What it says about this wiki’s evidence spectrum
The wiki’s standing caution is that vendor benchmark figures are first-party snapshots and need a neutral benchmark behind them (growth edge 2). This source moves the problem one level down: a neutral benchmark is only as good as its labels, and the most-cited labels in the field are wrong at a measurable rate. A T2 independent benchmark run on a 6%-mislabeled test set is not neutral in the way the edge assumes.
That does not retire the edge. It changes what closing it would require — a benchmark that states its label quality, which almost none do.
Related
data-cascades · training-data-quality · machine-learning · demo-to-production-gap · curtis-northcutt · synthesis