Research map / Label-efficient and robust learning / Noisy labels
Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning
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 NPN, which combines adaptive partial label learning (hard and soft disambiguation via candidate-label frequency) with negative learning using all non-candidate labels plus a consistency regularization term, to train robustly on noisy labels without a clean-sample prior.
Problem
Noisy label learning for image classification, where existing methods rely on prior assumptions such as a pre-defined drop rate or a small clean subset.
Approach
Decompose the label space adaptively into candidate labels (given label plus current top-1 prediction, tracked by a frequency set over epochs) for partial label learning, and use all remaining non-candidate labels as a complementary set for negative learning, with a weak-to-strong consistency loss for later training.
Key results
- On CIFAR100N Symmetric-80%, NPN-hard reaches 36.88% test accuracy, an 11.76-point improvement over the best listed baseline.
- On real-world datasets, NPN-hard achieves 86.02% (Web-Aircraft), 80.91% (Web-Bird), and 88.26% (Web-Car), surpassing Co-LDL by 4.05, 0.80, and 1.31 points.
- Ablation on CIFAR100N Sym-20%: adding NL raises accuracy from 35.50% to 54.70%, PLL adds 5.92 points, and consistency regularization adds 4.65 points to reach 65.27%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NPN-hard instead of the strongest baseline | CIFAR100N · CIFAR100N symmetric 80% noise | test accuracy: 27.41 -> 36.88 | Helped |
| Use NPN-hard instead of the strongest baseline | CIFAR100N · CIFAR100N asymmetric 50% noise | test accuracy: 30.50 -> 33.19 | Helped |
| Use NPN-soft instead of the strongest baseline | CIFAR100N · CIFAR100N asymmetric 50% noise | test accuracy: 30.50 -> 33.23 | Helped |
| Use NPN-hard instead of the strongest baseline | NPN · Web-Aircraft | test accuracy: 84.22 -> 86.02 | Helped |
| Use NPN-hard instead of the strongest baseline | NPN · Web-Bird | test accuracy: 80.11 -> 80.91 | Helped |
| Use NPN-hard instead of the strongest baseline | NPN · Web-Car | test accuracy: 86.95 -> 88.26 | Helped |
| Add negative learning to standard noisy-label training (ablation) | NPN · CIFAR100N symmetric 20% noise | test accuracy: 35.50 -> 54.70 | Helped |
| Add partial label learning to standard training with negative learning (ablation) | NPN · CIFAR100N symmetric 20% noise | test accuracy: 54.70 -> 60.62 | Helped |
| Add consistency regularization (ablation) | NPN-hard · CIFAR100N symmetric 20% noise | test accuracy: 60.62 -> 65.27 | Helped |
| Use hard rather than soft disambiguation (ablation) | NPN · CIFAR100N symmetric 20% noise | test accuracy: 62.76 -> 65.27 | Helped |
| Use hard rather than soft disambiguation (ablation) | NPN · CIFAR100N asymmetric 50% noise | test accuracy: 33.23 -> 33.19 | Hurt |
| Include more top-k predicted labels in the candidate set (ablation) | NPN · CIFAR100N | test accuracy | Hurt |
Surprising findings
- Selecting more labels into the candidate set makes disambiguation harder and lowers test accuracy, so the method keeps only the given label and the top-1 prediction.
Limitations
- Relies on the reliability of candidate and complementary labels, which can degrade in later training when the model overfits noisy samples.
- Hyper-parameter sensitivity to alpha and beta weights is studied only in a single setting.
Experimental setup
Datasets: CIFAR100N Web-Aircraft Web-Bird Web-Car
Benchmarks: CIFAR100N Web-Aircraft Web-Bird Web-Car
Metrics: test accuracy
Baselines: Standard Decoupling Co-teaching Co-teaching+ JoCoR Co-LDL SPRL AGCE PENCIL DivideMix AFM Self-adaptive
Builds on: NLNL JNPL Co-LDL DivideMix
Closest papers by meaning
- Partial Multi-Label Learning with Noisy Label Identification (AAAI 2020 · similarity 0.91)
- Complementary Label Learning with Positive Label Guessing and Negative Label Enhancement (ICLR 2025 · similarity 0.91)
- FREDIS: A Fusion Framework of Refinement and Disambiguation for Unreliable Partial Label Learning (ICML 2023 · similarity 0.90)
- Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label Configurations (NeurIPS 2024 · similarity 0.89)
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.89)
- Conformal Prediction for Partial Label Learning (AAAI 2025 · 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.