Research map / Label-efficient and robust learning / Noisy labels
Early Stopping Against Label Noise Without Validation Data
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 Label Wave, an early stopping method for learning with noisy labels that tracks prediction changes on the training set (no hold-out validation data) and stops at the first local minimum.
Problem
Early stopping under label noise usually requires a held-out validation set, which wastes training data and is itself unreliable when labels are noisy.
Approach
Counts the number of training examples whose predicted labels change between consecutive epochs ("prediction changes"), smooths it with a moving average, and stops at its first local minimum, which the authors observe coincides with the onset of fitting mislabeled examples and the test-error minimum.
Key results
- Label Wave selects models within about 0.2% to 0.9% of the global maximum test accuracy across the reported settings (e.g., 0.65% at 20% symmetric noise on CIFAR-10 with ResNet-18).
- On CIFAR-10 with 40% symmetric noise, Label Wave improves CE test accuracy to 81.61% vs 80.29% with 5% validation and 77.38% with 30% validation; SOP reaches 88.42% vs 86.40% with 5% validation.
- Label Wave attains the strongest negative Pearson correlation with test accuracy at k=3 (-0.9637) for the moving-average window.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Label Wave instead of 5% hold-out validation for stopping ELR | ELR on ResNet-18 · CIFAR-10, 40% symmetric noise | test accuracy: 90.23±0.47% -> 90.45±0.52% | Helped |
| Use Label Wave instead of 5% hold-out validation for stopping SOP | SOP on ResNet-34 · CIFAR-100, 40% symmetric noise | test accuracy: 67.27±0.45% -> 68.53±0.30% | Helped |
| Use Label Wave instead of 10% hold-out validation for stopping ELR | ELR on ResNet-18 · CIFAR-10, 40% instance-dependent noise | test accuracy: 87.68±0.32% -> 89.59±0.16% | Helped |
| Use Label Wave instead of 10% hold-out validation for stopping SOP | SOP on ResNet-34 · CIFAR-100, 40% instance-dependent noise | test accuracy: 67.83±0.15% -> 69.35±0.71% | Helped |
| Add batch normalization and data augmentation to Label Wave (ablation) | Label Wave on CIFAR-10 · CIFAR-10, 40% symmetric noise | test accuracy: 63.00±0.86% -> 66.79±0.39% | Helped |
| Add dropout to Label Wave + BN + DA (ablation) | Label Wave on CIFAR-10 · CIFAR-10, 40% symmetric noise | test accuracy: 81.61±0.44% -> 83.57±0.24% | Helped |
| Add Mixup to Label Wave + BN + DA (ablation) | Label Wave on CIFAR-10 · CIFAR-10, 40% symmetric noise | test accuracy: 81.61±0.44% -> 82.38±0.56% | Helped |
| Add Mixup to Label Wave + BN + DA + Dropout (ablation) | Label Wave on CIFAR-10 · CIFAR-10, 40% symmetric noise | test accuracy: 83.57±0.24% -> 83.67±0.45% | Helped |
| Vary the moving-average window k from 1 to 3 (ablation) | Label Wave · CIFAR-10, 40% symmetric noise | Pearson correlation coefficient: -0.8565 -> -0.9637 | Helped |
| Increase the moving-average window k from 3 to 5 (ablation) | Label Wave · CIFAR-10, 40% symmetric noise | Pearson correlation coefficient: -0.9637 -> -0.9367 | Hurt |
| Increase the moving-average window k from 3 to 10 (ablation) | Label Wave · CIFAR-10, 40% symmetric noise | Pearson correlation coefficient: -0.9637 -> -0.9406 | Hurt |
| Apply Label Wave rather than global-maximum oracle selection | Label Wave model selection · CIFAR-10, 40% symmetric noise | test accuracy: 84.08±0.10 -> 83.43±0.32 | Hurt |
Surprising findings
- The paper reports that Label Wave does not apply when label noise is very low or absent, and that robust regularization can remove the learning-confusion stage so the method may fail to find a stopping point.
Limitations
- Not applicable under very low or no label noise, or when robust regularization removes the learning-confusion stage.
- The transition at Point 2 (from confused to memorization) is not explained by the proposed metrics.
- Experiments are mostly on ResNet-18 with CIFAR-10 at 40% noise and single-setting hyperparameters.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N Clothing1M WebVision Food101 Tiny-ImageNet NEWS
Benchmarks: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N Clothing1M WebVision Food101 Tiny-ImageNet NEWS
Metrics: test accuracy difference from global maximum test accuracy Kendall tau correlation Pearson correlation coefficient
Baselines: hold-out validation (5%, 10%, 20%, 30%) global maximum test accuracy CE Taylor-CE ELR CDR CORES NLS SOP
Builds on: k-epoch learning Co-teaching ELR SOP
Closest papers by meaning
- Understanding and Improving Early Stopping for Learning with Noisy Labels (NeurIPS 2021 · similarity 0.91)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.88)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.86)
- Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks (ICLR 2024 · similarity 0.85)
- How Does Label Noise Gradient Descent Improve Generalization in the Low SNR Regime? (NeurIPS 2025 · similarity 0.85)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · 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.