Research map / Label-efficient and robust learning / Noisy labels
Enhancing Sample Selection Against Label Noise by Cutting Mislabeled Easy Examples
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Early Cutting, a recalibration step for sample selection in learning with noisy labels that removes "Mislabeled Easy Examples" (early-learned, confidently predicted, high-loss, low-gradient-norm noisy samples) using a later-epoch model.
Problem
Sample selection methods for learning with noisy labels rely on early-learned samples as confident, but some mislabeled examples learned early disproportionately harm generalization.
Approach
Identify Mislabeled Easy Examples (MEEs) among early-learned samples using a later-training-stage model via high loss, high confidence and low input-gradient norm, and remove them iteratively from the confident subset before final training.
Key results
- On CIFAR-100 with ResNet-34 and 40% symmetric noise, Early Cutting reaches 72.77% vs 67.68% for RLM (the strongest non-semi-supervised competitor in that table).
- On full ImageNet-1k with 40% symmetric noise, Early Cutting achieves 73.28% test accuracy vs 67.99% for cross-entropy and 71.42% for Late Stopping.
- With MixMatch semi-supervised extension (Early Cutting+), CIFAR-100 50% symmetric noise reaches 75.6%, above SFT+ (75.2%) and DivideMix (74.6%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply Early Cutting to dynamic-based selection | Dynamic-based sample selection · CIFAR-10, 40% instance-dependent noise | test accuracy: 83.12% -> 84.31% | Helped |
| add Early Cutting to iterative sample selection | CIFAR-10 sample selection · CIFAR-10, symmetric 20%, symmetric 40%, instance-dependent 20%, and instance-dependent 40% noise | test accuracy: 93.79 $\pm$ 0.14% vs 93.11 $\pm$ 0.29%; 91.80 $\pm$ 0.18% vs 91.06 $\pm$ 0.17%; 93.40 $\pm$ 0.22% vs 93.13 $\pm$ 0.05%; | Helped |
| add Early Cutting to iterative sample selection | CIFAR-100 sample selection · CIFAR-100, symmetric and instance-dependent noise at 20% and 40% | test accuracy: 76.20 $\pm$ 0.27% vs 73.90 $\pm$ 0.34%; 72.77 $\pm$ 0.17% vs 67.68 $\pm$ 0.36%; 75.03 $\pm$ 0.23% vs 72.47 $\pm$ 0.39%; | Helped |
| use Early Cutting rather than the strongest listed baseline | CIFAR-N sample selection · CIFAR-N 10N Random 1, 10N Random 2, 10N Random 3, 10N Worst, and 100N Fine | test accuracy: 92.50 $\pm$ 0.14% vs 92.21 $\pm$ 0.37%; 92.65 $\pm$ 0.11% vs 92.27 $\pm$ 0.31%; 92.36 $\pm$ 0.43% vs 92.07 $\pm$ 0.72%; | Helped |
| use Early Cutting rather than the strongest listed baseline | ResNet-50 sample selection · WebVision validation, ILSVRC12 validation, and ImageNet-1k symmetric 40% noise | test accuracy: 73.81% vs 72.32%; 71.20% vs 69.86%; 73.28% vs 71.42% | Helped |
| reduce the sample-selection iterations from three to one (ablation) | Early Cutting on CIFAR-100 · CIFAR-100, symmetric 40% noise | test accuracy: 69.94 $\pm$ 0.30 -> 69.53 \(\pm\) 0.10 | Hurt |
| add MEE filtering to the initially selected confident subset | CIFAR-10 sample selection · Representative ablation setting | test accuracy: 83.12% -> 84.57% | Helped |
| use an early cutting rate to focus MEE search on the earliest learned examples (ablation) | Early Cutting · Representative ablation setting | test accuracy: 84.12% -> 84.57% | Helped |
| add the loss-values criterion (ablation) | Early Cutting · Representative ablation setting | test accuracy: 82.36% -> 84.57% | Helped |
| add the confidence-scores criterion (ablation) | Early Cutting · Representative ablation setting | test accuracy: 84.10% -> 84.57% | Helped |
| add the gradient-norms criterion (ablation) | Early Cutting · Representative ablation setting | test accuracy: 84.07% -> 84.57% | Helped |
| vary the loss, confidence, gradient-norm thresholds, and early cutting rate over the tested ranges (ablation) | Early Cutting · CIFAR-10 and CIFAR-100, symmetric 40% noise | test accuracy: <1% | No clear effect |
Surprising findings
- Mislabeled examples learned earlier by the model harm generalization more than later-learned mislabeled examples, even though early-learned samples are usually trusted as clean.
- Using a later-stage (overfitting-onset) model to recalibrate the early-selected subset is counterintuitive but improves selection.
Limitations
- Validation confined to visual classification; applicability to NLP or tabular data untested.
- Justification of MEE harm is primarily observational without formal theoretical analysis.
- Iterative procedure adds computational overhead from input gradient computation at later stages; potential to filter minority-class clean hard examples under long-tailed or severe noise.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CIFAR-N WebVision ILSVRC12 ImageNet-1k
Benchmarks: CIFAR-10 CIFAR-100 CIFAR-N WebVision ILSVRC12 ImageNet-1k
Metrics: test accuracy
Baselines: Cross-Entropy GCE Co-teaching Me-Momentum Self-Filtering VOG Late Stopping Misdetect RLM Student Loss CSGN CORES^2
Builds on: Late Stopping Co-teaching MixMatch Learning speed / memorization effect
Closest papers by meaning
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.91)
- Debiased Sample Selection for Learning with Noisy Labels (CVPR 2026 · similarity 0.90)
- Enhancing Semi-Supervised Learning via Representative and Diverse Sample Selection (NeurIPS 2024 · similarity 0.86)
- Towards a statistical theory of data selection under weak supervision (ICLR 2024 · similarity 0.86)
- Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 2025 · similarity 0.85)
- Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.