Source-linked AI summary
Identifying Mislabeled Data using the Area Under the Margin Ranking
Geoff Pleiss, Tianyi Zhang, Ethan R. Elenberg, Kilian Q. Weinberger
TL;DR
Training sets can contain ambiguous or mislabeled samples that harm generalization, motivating automatic identification and removal. The paper uses the Area Under the Margin statistic and artificial threshold samples to separate mislabeled data from clean data during training. Across synthetic and real-world settings, the method improves classification performance, including a 1.2% test-error reduction after removing 13% of CIFAR100 data, while recall can struggle under extreme asymmetric noise.
Problem
Ambiguous and mislabeled training samples can harm generalization, while large datasets are costly to inspect manually, creating a need for automated identification and auditing.
Method
The method ranks samples with AUM, the average difference between assigned-class and highest non-assigned-class logits, and learns a threshold from purposefully mislabeled artificial-class samples.
Results
The method improves performance across benchmark and real-world settings; removing 13% of CIFAR100 reduces test error by 1.2% for ResNet-32.
Takeaways & Limitations
AUM provides a plug-and-play way to identify and remove suspected mislabeled data, while its ranking remains robust across architectures and threshold choices.
Takeaways & Limitations
Under systematic asymmetric noise, recall struggles at 30% and 40% noise, and the method is not competitive with the best approach at 40%.
Abstract
from arXiv · showhide
Not all data in a typical training set help with generalization; some samples can be overly ambiguous or outrightly mislabeled. This paper introduces a new method to identify such samples and mitigate their impact when training neural networks. At the heart of our algorithm is the Area Under the Margin (AUM) statistic, which exploits differences in the training dynamics of clean and mislabeled samples. A simple procedure - adding an extra class populated with purposefully mislabeled threshold samples - learns a AUM upper bound that isolates mislabeled data. This approach consistently improves upon prior work on synthetic and real-world datasets. On the WebVision50 classification task our method removes 17% of training data, yielding a 1.6% (absolute) improvement in test error. On CIFAR100 removing 13% of the data leads to a 1.2% drop in error.
1 Introduction
The paper targets harmful or ambiguous labels that limit generalization and proposes identifying them from training dynamics, then removing them using AUM-based thresholding. The method is designed to be plug-and-play and improves classification performance on benchmark data.
- Motivation: Mislabeled and ambiguous samples can limit generalization, even in highly curated datasets and despite overparameterized networks fitting noisy labels.The paper motivates automated dataset auditing because large datasets are costly to inspect manually.
- Motivation: The goal is to automatically identify and remove mislabeled samples so practitioners can reduce memorization, improve generalization, and audit annotation errors.The method is intended to achieve high precision and recall without requiring a multi-stage pipeline.
- Method: Training dynamics expose conflicting updates when an image’s assigned label disagrees with its hidden true class, producing informative logit trajectories.For example, a bird mislabeled as a dog receives opposing signals from its own label and from other training examples.
- Method: AUM measures the average difference between the assigned-class logit and the largest non-assigned-class logit, with mislabeled samples receiving lower values.The method adds an artificial class containing purposefully mislabeled threshold samples and uses their AUM statistics as a separation threshold.
- Results: The method is compatible with any classification network and requires logging training logits; removing identified samples improves benchmark performance.On CIFAR100, removing 13% of the dataset reduced test error by 1.2% for ResNet-32.
2 Related Work
The paper positions its approach among methods for learning with noisy data, including robust losses, label correction, filtering, auxiliary networks, and semi-supervised reuse. Unlike methods focused on reusing untrusted data, it emphasizes identifying and discarding mislabeled samples with minimal training changes.
- Existing approaches: Prior noisy-label research includes new architectures, curriculum learning, label correction, robust loss functions, and theoretical guarantees for selected regularization schemes.The paper aims to improve robustness while also increasing training-set quality.
- Existing approaches: Related deep-learning methods identify mislabeled data through cross-validation, influence functions, auxiliary networks, or signals extracted from training.The paper shares a pipeline that ranks examples and applies a learned threshold before classifier training.
- Scope distinction: Noisy-data methods also combine trusted subsets with semi-supervised learning, pseudo-labeling, or MixUp to reuse remaining untrusted data.This paper instead focuses primarily on identifying correctly labeled data and discards samples flagged as mislabeled.
3 Identifying Mislabeled Data
The method identifies mislabeled training samples by ranking examples with AUM, then learns a dataset-specific cutoff from artificial threshold samples and removes examples below it.
- Mislabeled samples are defined as examples whose assigned labels do not match their inputs, whereas correctly labeled samples may be easy or hard to learn.
- AUM averages each sample’s margin across training epochs, where the margin compares its assigned-class logit with the highest non-assigned-class logit.The margin is negative for incorrect predictions and positive for confident correct predictions.
- The method expects mislabeled examples to have smaller margins and therefore lower AUMs than correctly labeled examples because updates from similar samples oppose their assigned labels.
- Threshold samples are created by assigning a subset of training examples to a nonexistent extra class, making them mimic mislabeled data while minimally affecting the primary classification task.The extra class is added to the network output, and its samples require memorization to raise the assigned logit.
- On noisy CIFAR10/100, the 99th-percentile threshold AUM cleanly separates correctly labeled and mislabeled samples, although the AUM threshold is dataset dependent.
- The procedure trains with threshold samples, computes the 99th-percentile threshold AUM, and identifies non-threshold training examples whose AUM is at or below that value.Training stops before the first learning-rate drop to limit memorization of difficult or mislabeled examples; a second run identifies remaining mislabeled samples.
4 Experiments
Experiments evaluate AUM for identifying mislabeled samples and improving models after removing them, across synthetic and real-world datasets. Results show strong identification and generally improved or preserved test performance, with reduced effectiveness under highly systematic asymmetric noise.
- Mislabeled Sample Identification: AUM achieved the highest precision and recall in most synthetic-noise settings, consistently reaching at least 90% for both metrics at high noise without supervision or a known noise model.The most challenging cases were CIFAR100 and Tiny ImageNet with low noise.
- Robust Training on Synthetic Noisy Datasets: On Tiny ImageNet, AUM matched or surpassed oracle test performance in most uniform-noise settings.The comparison used ResNet-32 models and included Data Parameters, DY-Bootstrap, and INCV.
- Real-World Datasets: AUM cleaning reduced WebVision50 error from 21.4% to 19.8% after flagging 17.8% of the data as mislabeled.DY-Bootstrap did not reduce error over standard training, while INCV removed more data and produced higher WebVision50 error.
- Real-World Datasets: AUM reduced full Clothing1M error from 33.5% to 29.6%, while removing 10.7% of the data.The method removed fewer samples on the full dataset than on the 100K subset, possibly because larger datasets require more training to memorize.
- Mostly-clean Datasets: On mostly-clean datasets, AUM reduced CIFAR100 error from 33.0% to 31.8% while removing 13% of the data, and changed ImageNet top-1 error from 24.2% to 24.4% after removing 2%.The procedure removed 3% of CIFAR10 and 24% of Tiny ImageNet data.
- AUM Ranking versus Margin Ranking: AUM’s running average produced a more stable and separable signal than un-averaged margin values, whose clean and noisy trajectories overlapped during training.The averaging improves the signal-to-noise ratio of margin trajectories.
- Effect of Data Augmentation: With 40% CIFAR10 noise, AUM reduced error from 43% to 12% with augmentation and from 51% to 20% without augmentation.These results show effectiveness under both augmentation settings.
- Robustness to Architecture and Hyperparameters: AUM ranking was robust across architectures, achieving 98% Spearman correlation, while changing the threshold percentile to 90% altered final test error by only 4% relatively.The authors interpret the ranking as capturing dataset-dependent rather than model-dependent properties.
5 Discussion and Conclusion
The paper presents AUM with threshold samples as a simple approach for identifying noisy labels and reports practical usability across classifiers and training workflows.
- AUM can be combined with data augmentation and semi-supervised learning methods for noisy-data training.
- The approach improves classification performance after removing identified mislabeled samples, including on relatively clean and very noisy real-world datasets.
- The method works with any classifier without architecture or training-procedure changes, requiring only logged logits and one additional training round.
- The baseline comparisons exclude Co-Teaching because it requires a noise estimate, unlike the evaluated methods.
Broader Impact
The method could support both the use of weakly labeled data and the auditing of established datasets, but automated curation may introduce risks in sensitive settings.
- Identifying harmful training examples could enable broader use of cheap weakly labeled data and help audit datasets such as ImageNet.
- Automated identification may create or amplify dataset biases, while auditing and curation can have unintended consequences for security- or privacy-sensitive applications.
- Because standard pretraining datasets support many downstream applications, improving their quality could affect numerous tasks.
Supplementary Information for: Identifying Mislabeled Data using the Area Under the Margin
The supplementary information identifies the paper’s authors and their institutional affiliations.
- Tianyi Zhang is affiliated with Stanford University.
- Kilian Q. Weinberger is affiliated with ASAPP and Cornell University.
A Experiment Details
Experiments use specified PyTorch training protocols across several image datasets, with AUM computed during an earlier training phase and iteration counts preserved after data removal.
- Experiments use PyTorch, omit early stopping without trusted validation data, and report final-epoch test errors.
- Tables generally report means and standard deviations over four random-seed trials, while larger datasets use a single trial.
- CIFAR10, CIFAR100, and Tiny ImageNet models use ResNet-32 training for 300 epochs with specified SGD, learning-rate, regularization, batch-size, and augmentation settings.
- AUM computation for smaller datasets runs until the first learning-rate drop, using 150 epochs and batch size 64 to increase SGD variance.
- After removing data, batch sizes are adjusted to preserve the original number of training iterations.
- WebVision50 and Clothing100K use ResNet-50 models trained from scratch for 180 epochs with standard augmentation and scheduled learning-rate drops.
- The ImageNet procedure uses 90 epochs, learning-rate drops at epochs 30 and 60, and computes AUM through epoch 30.
B Additional Ablation Studies
AUM rankings remain highly consistent across network architectures and are robust to the threshold-sample percentile used for cleaning, outperforming alternative training-based metrics in consistency.
- Consistency across architectures: > 98% correlation holds between AUM rankings from ResNet and DenseNet models across depths.The comparison uses CIFAR10 with 40% label noise.
- Consistency across architectures: AUM rankings are more consistent than training loss at 75% correlation and validation loss at 40% correlation.The reported comparisons use the same CIFAR10 noisy-data setting across network architectures.
- Consistency across architectures: 75% correlation is achieved by the margin without a running average, compared with AUM’s higher cross-network consistency.AUM’s running average across all epochs contributes to its ranking consistency.
- Robustness against Threshold Sample Percentile: 39% test error at the 99th-percentile threshold versus 41% at the 90th percentile under 40% label noise.On unmodified CIFAR100, final test error is virtually unaffected by the threshold percentile; with 40% noise, higher percentiles typically perform better.
- Robustness against Threshold Sample Percentile: The threshold-sample percentile has a relatively limited effect on AUM-cleaned model performance.The paper suggests using the 99th percentile, while reporting robustness across percentile choices.
C More Results for Real-World Datasets
On real-world datasets, AUM distributions need not be bimodal, yet dataset-specific threshold samples and worst-ranked images provide practical signals for inspecting potentially harmful data.
- AUM values: Real-world datasets do not exhibit the bimodal AUM behavior seen in synthetic mislabeled datasets.Figure S3 shows empirical AUM densities for the real-world datasets.
- AUM values: The 99% threshold sample differs across all real-world datasets.Figure S3 represents each threshold sample with a gray line.
- Example removed images: Worst-ranked images are shown for CIFAR10, CIFAR100, WebVision50, and Clothing1M.Figures S4–S6 provide the dataset-specific image examples.