PaperlayerView exact revision on arXiv

Source-linked AI summary

Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks

Curtis G. Northcutt, Anish Athalye, Jonas Mueller

arXiv:2103.14749v4stat.MLcs.AIcs.LG

TL;DR

Benchmark test sets may contain naturally occurring label errors, but their prevalence and consequences have been less studied than training-set noise. The paper identifies and human-validates errors across 10 datasets, then examines benchmark stability under increasing test-label noise. It finds that correcting labels can make lower-capacity models preferable, while the identified errors and noise estimates remain conservative because confident-learning failures and unflagged errors are not fully captured.

  • Problem

    Test-set label errors are less studied than training-set noise, yet they can destabilize benchmarks and mislead conclusions about which models perform best in real-world datasets.

  • Method

    The study applies confident learning to identify putative errors across 10 benchmark datasets and validates candidates through human evaluation before analyzing benchmark stability.

  • Results

    Corrected-label evaluation shows that higher-capacity models can perform worse on originally mislabeled examples, making lower-capacity models practically preferable as test-label noise increases.

  • Takeaways & Limitations

    Practitioners should distinguish corrected from original test accuracy and consider correcting test labels when selecting models for noisy real-world applications.

  • Takeaways & Limitations

    The reported noise estimates are lower bounds because confident-learning failures and label errors among non-flagged data are not fully accounted for.

Abstract

from arXiv · show

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.

1 Introduction

Test-set label errors are widespread but under-studied, threatening benchmark stability and potentially misleading model selection. The study characterizes these errors across 10 datasets and finds that correcting labels can reverse practical preferences between models.

  • Motivation: Test sets can contain errors that destabilize machine-learning benchmarks, despite often being treated as correct when drawn from the training distribution.This issue differs from the more commonly studied problem of noisy training labels.
  • Motivation: Researchers use benchmark test datasets to measure progress, while practitioners use noisier real-world datasets to choose models.Errors in those test sets could lead to incorrect conclusions about which models perform best in deployment.
  • Study scope: 10 datasets across computer vision, natural language processing, and audio processing were systematically studied using confident learning and human evaluation.The study estimates a lower bound of 3.3% errors on average, including 6% in the ImageNet validation set and over 5 million errors in QuickDraw.
  • Benchmark consequences: Higher-capacity models better reproduce noisy test labels but perform worse than lower-capacity models on corrected labels for originally mislabeled examples.This benchmark instability increases with the prevalence of mislabeled test data and is not traditional overfitting.
  • Benchmark consequences: A 6% increase in originally mislabeled ImageNet test examples makes ResNet-18 outperform ResNet-50 with corrected labels.The analysis incrementally increases noise prevalence to identify when benchmark rankings change.
  • Implications: Practitioners may benefit from correcting test labels and considering simpler models when real-world datasets are noisier than benchmark datasets.The paper recommends distinguishing corrected from original test accuracy when making deployment decisions.

2 Background and related work

Prior work largely studies noisy training labels or synthetic noise, while test-set label errors remain less examined. The paper uses confident learning to identify naturally occurring, class-conditional errors and validates their correctable nature with crowdsourcing.

  • Prior work: Research on noisy labels usually focuses on training data, often using synthetic noise or assuming robustness to label noise.These choices do not directly characterize naturally occurring errors in benchmark test sets.
  • Validation challenge: Crowd-sourced label validation is common but can be exorbitantly expensive for large datasets.The paper therefore prioritizes likely errors before human verification.
  • Approach: Confident learning is used to estimate label errors robustly across class-imbalanced datasets with asymmetric, non-uniform noise.The framework is model-agnostic and quantifies class-conditional label noise.
  • Validation findings: Crowdsourced validation confirms that benchmark noise is primarily systematic mislabeling rather than random noise or lack of signal.This validation targets the correctable nature of the identified errors.

3 Identifying label errors in benchmark datasets

The pipeline estimates class-conditional label noise with confident learning, ranks likely errors from out-of-sample predictions, and validates prioritized candidates rather than checking every example manually.

  • Noise characterization: Confident learning estimates the joint distribution of observed noisy labels and unknown true labels for a dataset with n examples and m classes.The method assumes label noise depends on the latent true class rather than directly on the data.
  • Noise characterization: The diagonal of the estimated joint-distribution matrix gives the probability that examples in each class are correctly labeled.The total label-error fraction is derived from the estimated correctly labeled mass.
  • Candidate ranking: Potential errors are identified automatically by ranking examples with a normalized margin between the given-label probability and the strongest alternative-label probability.The top estimated-error count is selected according to the estimated error fraction.
  • Prediction inputs: Out-of-sample predicted probabilities are obtained by pre-training and fine-tuning with cross-validation, or by using open-sourced pre-trained models when available.This produces the prediction inputs required for confident-learning noise characterization.
  • Dataset results: Automatic candidate identification avoids manually checking the entire test set, while Table 1 reports error rates ranging from 0.15% for MNIST to 10.12% for QuickDraw.The study uses dataset-specific test-set coverage notes for datasets without explicit public test sets.

4 Validating label errors with Mechanical Turk

The study human-validates confident-learning candidates through a Mechanical Turk protocol, using multiple workers and explicit label-agreement criteria. It also documents challenging false-positive cases and acknowledges that corrected labels remain imperfect.

  • The study sampled errors in QuickDraw and Amazon Reviews and checked all identified errors in the remaining datasets.
  • Workers classified whether they saw the given label, the CL-predicted label, both labels, or neither label.The interface also showed high-confidence examples of the given and predicted classes.
  • Each CL-flagged label error was independently presented to five workers for validation.
  • A label was validated as erroneous when fewer than three of five workers agreed that the data point had its given label.Otherwise, the original label was treated as correct; validated errors were further categorized, including correctable and multi-label cases.
  • Failure modes of confident learning: Confident learning can flag difficult examples that human validation finds correctly labeled, including cropped, ambiguous, or unusual images.Examples include a cropped sewing machine and an airplane viewed from the cockpit.
  • Failure modes of confident learning: The corrected test sets are improved but not 100% perfect, even when all five reviewers agree on a label correction.The paper reports that these failure-mode cases are rare.

5 Implications of label errors in test data

Test-set label errors can reverse model rankings when evaluation distinguishes original labels from corrected labels, making benchmark-based deployment decisions unstable as noise prevalence rises.

  • Original accuracy measures predictions against dataset labels, whereas corrected accuracy evaluates labels corrected by humans or removes ambiguous examples.
  • 5.1 Benchmarking on the correctable set: ResNet-18 significantly outperforms NASNet on corrected accuracy over the correctable set despite much worse original test accuracy.
  • 5.1 Benchmarking on the correctable set: NASNet-large drops from rank 1/34 to 29/34, while ResNet-18 rises from 34/34 to 1/34 after switching from original to corrected labels.
  • 5.1 Benchmarking on the correctable set: Lower-capacity models may resist learning systematic label-error patterns, while larger models can overfit benchmark-specific annotation quirks.
  • 5.2 Benchmark instability: On ImageNet, ResNet-18 overtakes ResNet-50 when originally mislabeled test examples increase by around 6% relative to the original pruned data.
  • 5.2 Benchmark instability: The analysis is a loose lower bound because unaccounted errors in non-CL-flagged data likely make corrected-versus-original accuracy gaps larger.

6 Expert review of CL-flagged and non-CL-flagged label errors in ImageNet

Expert review tested whether confident-learning flags identify more errors than random non-flagged examples and found that flagged examples are substantially more error-prone, while the broader ImageNet error rate may be much higher.

  • Experts reviewed randomly selected CL-flagged and non-CL-flagged ImageNet examples across classes to compare label-issue rates.
  • An image flagged by confident learning was 2.6x as likely to be erroneously labeled as a non-CL-flagged image.
  • MTurk results broadly agreed with expert review, supporting confident learning followed by crowdsourced verification for characterizing label errors.
  • ImageNet validation may contain closer to 20% label errors, up from the previously reported 6%, after accounting for errors missed among non-CL-flagged images.

7 Discussion

The paper finds widespread label errors in popular machine-learning benchmark test sets and argues that these errors can destabilize model comparisons. It recommends evaluating models on correctly labeled test sets, particularly for noisier real-world data.

  • Label errors are ubiquitous across test sets of popular machine-learning benchmarks.
  • Small increases in mislabeled test data can destabilize benchmark rankings.The authors report that lower-capacity models may outperform higher-capacity models in noisy real-world applications.
  • Model deployment decisions based solely on test accuracy warrant caution.The paper emphasizes judging models over correctly labeled test sets, especially for real-world datasets likely to be noisier than popular benchmarks.
  • The study evaluates 10 open-source datasets spanning computer vision, natural language, sentiment analysis, and audio.

A.1 Dataset details

The study examines label-generation procedures and potential error sources across 10 datasets in visual, text, sentiment, and audio modalities. It also describes the crowdsourced validation setup and illustrative cases of ambiguous or incorrect labels.

  • Label errors can arise from ambiguous handwriting, incomplete drawings, scraped images, or human labeling mistakes.
  • Text-dataset labels were derived from newsgroup membership, review sentiment scores, or Amazon star ratings.The study removes 2-star and 4-star Amazon reviews because of ambiguity with 1-star and 5-star reviews.
  • AudioSet labels were assigned through majority agreement among three human labelers, yet spot checks found errors from confusing labels, human error, and faint audio events.
  • Mechanical Turk workers were paid $7.20 per hour, with $1623.29 spent on human verification experiments.The authors state that results would likely improve with a larger budget.
  • The worker interface compared each candidate's given label with a confidently predicted class, while an example case showed that both labels could be wrong.

C Details of confident learning (CL) for finding label errors

Confident learning estimates an observed-label/true-label joint distribution to flag likely label errors for human review. It uses out-of-sample predicted probabilities and per-class thresholds to assign examples to confident bins, then normalizes the resulting counts.

  • Confident learning estimates an unnormalized confident joint between noisy observed labels and presumed true labels.
  • Examples are binned when their out-of-sample predicted probability for a class exceeds that class's threshold.The threshold is defined per class and is intended to provide robustness to heterogeneous class distributions.
  • When an example qualifies for multiple bins, it is counted only in the bin corresponding to its highest predicted probability.
  • The noisy-label transition estimate is obtained by normalizing the confident joint.The numerator calibrates row sums to the observed noisy-label prior, and the denominator makes the distribution sum to 1.

D Failure modes of confident learning

Confident learning can misestimate the sets of examples associated with noisy and true labels when predicted probabilities are inaccurate. The paper describes false exclusions and false inclusions caused by threshold-crossing errors.

  • Confident learning can fail to exactly estimate examples with noisy label i and actual label j.
  • A candidate may be falsely assigned to another class k when its predicted probability exceeds t_k even though k ≠ j.
  • Noisy predicted probabilities are modeled as ideal probabilities plus an error or deviation term.
  • The ideal predicted probability need not be 0 or 1 because some classes are mislabeled as others.
  • Failure Case 1 excludes a truly j-labeled example, whereas Case 2 incorrectly includes it under another class k.
  • Either failure case produces an imperfect estimate of label errors before human validation.Figure 2 presents uniquely challenging examples associated with potentially erroneous predicted probabilities.

E Reproducibility and computational requirements

The authors open-source a script that reproduces label errors across all 10 datasets, requiring about 5 minutes on a modern computer.

  • About 5 minutes are required to reproduce label errors for all 10 datasets using the open-source tutorial.The script is designed to run on a modern computer.

F Additional findings on implications of label errors in test data

Correcting test-set labels can substantially change benchmark rankings, with lower-capacity models sometimes outperforming higher-capacity models on ImageNet and CIFAR-10. These findings motivate evaluating generalization on correctly labeled test data and using hidden correctable errors to detect test-set overfitting.

  • ImageNet ranking stability: Benchmark rankings on ImageNet’s correctable subset change significantly when erroneous labels are replaced with corrected labels, including under a stricter agreement threshold.The stricter threshold requires all 5 of 5 human raters to independently select the same alternative label.
  • ImageNet ranking stability: Top-5 benchmark rankings remain unchanged after removing label errors, but rankings on the correctable subset change drastically between original and corrected labels.The corrected subsets contain 1428, 960, or 468 examples depending on the setting.
  • Benchmarking implications: The ranking changes may reflect overfitting during hyper-parameter tuning or to training-set noise, suggesting that secret correctable test examples could help detect test-set overfitting.This framework is proposed as a route toward more reliable benchmarking of generalization accuracy.
  • Capacity and corrected labels: Lower-capacity models tend to outperform higher-capacity models when ImageNet and CIFAR-10 models are benchmarked using corrected labels.The CIFAR-10 replication reports the same directional pattern as ImageNet, although its corrected-example count is limited.
  • Benchmarking implications: Corrected test data should measure predictions against error-free outputs rather than replicate noisy training distributions in real-world settings.The paper supports this conclusion with independent ImageNet and CIFAR-10 experiments.
  • Validation process: Expert review included 1934 ImageNet images, with each image reviewed by at least 2 experts, while reviewers were blinded to confident-learning flags.The review process included 934 CL-flagged and 1000 non-CL-flagged images and used consensus discussion for disagreements.
Loading 2103.14749v4…