[Paper Review] Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks

Curtis G. Northcutt, Anish Athalye+1 more|arXiv (Cornell University)|Mar 26, 2021

Machine Learning and Data Classification49 references103 citations

TL;DR

The paper shows pervasive label errors in test sets of 10 common benchmarks, demonstrates their impact on benchmark stability, and argues for correcting test labels to better reflect real-world model performance, especially for noisier data.

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 at least 3.3% errors across the 10 datasets, where for example label errors comprise at least 6% of the ImageNet validation set. Putative label errors are identified using confident learning algorithms and then human-validated via crowdsourcing (51% of the algorithmically-flagged candidates are indeed erroneously labeled, on average across the datasets). 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 ResNet-50 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%. Test set errors across the 10 datasets can be viewed at https://labelerrors.com and all label errors can be reproduced by https://github.com/cleanlab/label-errors.

Motivation & Objective

• Quantify the prevalence of label errors in test sets across 10 benchmark datasets in vision, language, and audio.
• Evaluate how test-set label errors affect benchmark stability and model selection.
• Provide open-source resources and a framework to assess corrected vs. original test accuracy for practical deployment insights.

Proposed method

• Apply confident learning (CL) to estimate the joint distribution of observed vs. true labels in test sets.
• Prioritize likely label errors for human validation via Mechanical Turk to obtain corrected labels.
• Use expert review to validate CL-flagged and non-CL-flagged errors in ImageNet and quantify error prevalence.
• Analyze how corrected vs. original test accuracy affects model rankings on ImageNet and CIFAR-10.
• Incrementally vary noise prevalence in test sets to study benchmark instability and model selection sensitivity.

Experimental results

Research questions

• RQ1 What is the prevalence of label errors in test sets of widely used ML benchmarks across modalities?
• RQ2 How do test-set label errors influence benchmark stability and the perceived superiority of higher-capacity models?
• RQ3 Do corrected test labels change model rankings, and under what conditions might smaller models outperform larger ones in practice?
• RQ4 Can we provide open resources to reproduce label corrections and assess performance under corrected labels?

Key findings

• Average test-set label error rate is at least 3.3% across 10 datasets.
• ImageNet validation set contains 6% putative label errors; QuickDraw contains over 10% errors (interpreted from CL-categorized results).
• Human validation via MTurk shows about 51% of algorithm-flagged candidates are actually mislabeled on average across datasets.
• Benchmark rankings largely remain the same when using corrected labels, but higher-capacity models disproportionately reflect noise, increasing instability as mislabeled data prevalence grows.
• On ImageNet, ResNet-18 can outperform ResNet-50 when the prevalence of originally mislabeled examples increases by about 6%; on CIFAR-10, VGG-11 can outperform VGG-19 when mislabeled prevalence increases by about 5%.
• Corrected accuracy (A*) on the correctable subset can reverse many apparent gains observed with original accuracy (Ã) on noisy test data, indicating practical deployment implications.