Research map / Label-efficient and robust learning / Noisy labels
Mitigating Label Noise through Data Ambiguation
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 Robust Data Ambiguation (RDA), a loss that augments training targets with credal sets built from confidence-thresholded possibility distributions, so the learner keeps the noisy label but adds model-predicted alternatives, robust to label noise without extra parameters.
Problem
Deep networks trained with cross-entropy memorize corrupted training labels, hurting generalization; existing robust losses lack flexibility and label-correction methods add complexity.
Approach
Instead of committing to one label or replacing it, ambiguate the target by keeping the observed label and adding model-predicted candidates whose confidence exceeds a threshold, forming a credal set; the superset optimistic loss with label relaxation then yields a convex robust loss with no extra parameters and CE-level cost.
Key results
- On CIFAR-10 with 50% symmetric noise, RDA reaches 86.47% test accuracy versus 83.36% for GCE and 84.18% for NCE+AUL among methods without extra parameters.
- On CIFAR-10N (Worst split), RDA achieves 82.91% versus 81.28% for the best parameter-free baseline NCE+AUL.
- On WebVision and Clothing1M with ResNet50, RDA reports 70.23% and 71.42% test accuracy, slightly above the listed baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RDA instead of the strongest baseline SOP | ResNet34 · CIFAR-10 symmetric noise 25% | accuracy: 92.58 ± 0.08 -> 91.48 ± 0.22 | Hurt |
| Use RDA instead of the strongest baseline SOP | ResNet34 · CIFAR-10 symmetric noise 50% | accuracy: 89.21 ± 0.33 -> 86.47 ± 0.42 | Hurt |
| Use RDA instead of the strongest baseline SOP | ResNet34 · CIFAR-10 symmetric noise 75% | accuracy: 76.16 ± 4.88 -> 48.11 ± 15.41 | Hurt |
| Use RDA instead of the strongest baseline ELR | ResNet34 · CIFAR-100 symmetric noise 25% | accuracy: 73.66 ± 1.87 -> 70.03 ± 0.32 | Hurt |
| Use RDA instead of the strongest baseline SOP | ResNet34 · CIFAR-100 symmetric noise 50% | accuracy: 64.28 ± 1.44 -> 59.83 ± 1.15 | Hurt |
| Use RDA instead of the strongest baseline SOP | ResNet34 · CIFAR-100 symmetric noise 75% | accuracy: 40.59 ± 1.62 -> 26.75 ± 8.83 | Hurt |
| Use RDA instead of the strongest baseline ELR | ResNet34 · CIFAR-10N Random 1 | accuracy: 91.35 ± 0.29 -> 90.43 ± 0.03 | Hurt |
| Use RDA instead of the strongest baseline ELR | ResNet34 · CIFAR-10N Aggregate | accuracy: 92.68 ± 0.03 -> 91.71 ± 0.38 | Hurt |
| Use RDA instead of the strongest baseline ELR | ResNet34 · CIFAR-100N Noisy | accuracy: 62.80 ± 0.27 -> 59.22 ± 0.26 | Hurt |
| Use RDA instead of AGCE | ResNet50 · WebVision | accuracy: 69.4 -> 70.23 | Helped |
| Use RDA instead of GCE | ResNet50 · Clothing1M | accuracy: 69.75 -> 71.42 | Helped |
Surprising findings
- The method becomes less effective as noise level increases, e.g., at 75% symmetric noise on CIFAR-10 it is not competitive with the best methods (48.11% vs 76.16% for SOP).
- ELR and SOP, which use extra per-instance parameters, outperform RDA in most synthetic-noise settings.
Limitations
- Effectiveness decreases with increasing noise level (e.g., 75% symmetric noise).
- Requires tuning of threshold beta and relaxation alpha, with beta schedule chosen heuristically.
- Gains on large-scale WebVision and Clothing1M are modest.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 WebVision Clothing1M CIFAR-10N CIFAR-100N
Benchmarks: CIFAR-10 CIFAR-100 CIFAR-10N CIFAR-100N WebVision Clothing1M
Metrics: test accuracy
Baselines: CE label smoothing label relaxation GCE NCE NCE+AGCE NCE+AUL CORES ELR SOP
Builds on: label relaxation superset learning optimistic superset loss generalized cross-entropy ELR
Closest papers by meaning
- Robust Classification via Regression for Learning with Noisy Labels (ICLR 2024 · similarity 0.91)
- Two Wrongs Don’t Make a Right: Combating Confirmation Bias in Learning with Label Noise (AAAI 2023 · similarity 0.90)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.90)
- Error-Bounded Correction of Noisy Labels (ICML 2020 · similarity 0.89)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.88)
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.88)
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.