Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks
Curtis G. Northcutt ChipBrain, MIT, Cleanlab、Anish Athalye MIT, Cleanlab、Jonas Mueller AWS Correspondence to: or
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 original
中文速览
十大主流机器学习基准数据集（涵盖图像、文本、音频领域）的测试集里，其实藏着大量被错误标注的样本。研究者用"置信学习"（confident learning）算法自动筛查可疑错误，再通过众包平台让真人验证，发现平均至少3.3%的测试样本标注有误，仅ImageNet验证集就有约6%（近3000张）的错误标签。更出乎意料的是，这些标注噪声会扭曲模型选拔结果：大模型（如NASNet、ResNet-50）更擅长"拟合"那些错误标签，在原始噪声测试集上看起来更强，但一旦用修正后的标签重新评估，它们反而不如小模型（如ResNet-18）——ImageNet上只需噪声比例再提高6%，ResNet-18就能反超ResNet-50。这项研究警示我们：在真实世界的嘈杂数据场景中，仅凭标准测试集准确率选模型可能得出错误结论，先纠正测试集标注、再做评估，才是更可靠的做法。
原文 arXiv:2103.14749;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2103.14749v4