Research map / Label-efficient and robust learning / Noisy labels
Pi-DUAL: Using privileged information to distinguish clean from noisy labels
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 Pi-DUAL, a dual-path network that uses privileged information to route clean and noisy labels to a prediction network and a noise network via a PI-driven gate, improving noisy-label test accuracy and wrong-label detection.
Problem
Learning with label noise when privileged information (PI) is available at training time but not at test time; existing PI methods often fail to outperform no-PI baselines in preventing overfitting to noisy labels.
Approach
Decompose output logits into a prediction network over input features x and a noise network over PI a, combined by a sigmoid gate computed from a; the gate routes clean vs wrong labels so the prediction network is not forced to fit noise. At inference only the prediction network is used.
Key results
- Pi-DUAL reaches 71.3% test accuracy on CIFAR-10H (worst) vs 64.9% for TRAM, a +4.5-point gain over the best PI baseline.
- On ImageNet-PI high-noise, Pi-DUAL achieves 62.1% vs 55.3% for AFM (+6.8 points).
- Confidence thresholding of the prediction network gives wrong-label detection AUC of 0.954 on CIFAR-10H and 0.986 on ImageNet-PI high-noise, outperforming baselines on most datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Pi-DUAL instead of TRAM++ | CIFAR-10H (worst) · CIFAR-10H (worst) | test accuracy: 66.8 -> 71.3 | Helped |
| Use Pi-DUAL instead of the strongest baseline | CIFAR-10N (worst) · CIFAR-10N (worst) | test accuracy: 86.6 -> 84.9 | Hurt |
| Use Pi-DUAL instead of the strongest baseline | CIFAR-100N (fine) · CIFAR-100N (fine) | test accuracy: 64.0 -> 64.2 | Helped |
| Use Pi-DUAL instead of AFM | ImageNet-PI (low-noise) · ImageNet-PI (low-noise) | test accuracy: 70.3 -> 71.6 | Helped |
| Use Pi-DUAL instead of AFM | ImageNet-PI (high-noise) · ImageNet-PI (high-noise) | test accuracy: 55.3 -> 62.1 | Helped |
| Add a gating network (ablation) | Pi-DUAL · CIFAR-10H, CIFAR-10N, CIFAR-100N, ImageNet-PI low-noise and high-noise | test accuracy: 61.5 -> 71.3; 84.5 -> 84.9; 59.0 -> 64.2; 67.9 -> 71.6; 47.8 -> 62.1 | Helped |
| Add a noise network (ablation) | Pi-DUAL · CIFAR-10H, CIFAR-10N, CIFAR-100N, ImageNet-PI low-noise and high-noise | test accuracy: 59.7 -> 71.3; 82.4 -> 84.9; 59.7 -> 64.2; 71.6 -> 71.6; 62.3 -> 62.1 | Mixed |
| Keep gating in logit space rather than probability space (ablation) | Pi-DUAL · CIFAR-10H, CIFAR-10N, CIFAR-100N, ImageNet-PI low-noise and high-noise | test accuracy: 62.2 -> 71.3; 81.6 -> 84.9; 59.4 -> 64.2; 71.0 -> 71.6; 60.4 -> 62.1 | Helped |
| Add dataset-specific PI features alongside random sample identifiers (ablation) | Pi-DUAL · CIFAR-10H, CIFAR-10N, CIFAR-100N, ImageNet-PI low-noise and high-noise | test accuracy: 53.5 -> 71.3; 83.7 -> 84.9; 61.8 -> 64.2; 68.4 -> 71.6; 47.0 -> 62.1 | Helped |
| Use Pi-DUAL confidence thresholding rather than the strongest listed baseline | Noise detection · CIFAR-10H (worst) | AUC: 0.834 -> 0.954 | Helped |
| Use Pi-DUAL confidence thresholding rather than the strongest listed baseline | Noise detection · CIFAR-10N (worst) | AUC: 0.968 -> 0.962 | No clear effect |
| Use Pi-DUAL confidence thresholding rather than the strongest listed baseline | Noise detection · CIFAR-100N (fine) | AUC: 0.889 -> 0.911 | Helped |
Surprising findings
- On ImageNet-PI, the noise network fits almost no samples, and removing it barely changes accuracy; the gate alone appears sufficient.
- Gate thresholding degrades substantially on CIFAR-10N and CIFAR-100N, where PI is aggregated and low-resolution.
Limitations
- Performance depends heavily on the quality of PI; gains are small or absent on datasets with low-resolution PI (CIFAR-10N, CIFAR-100N).
- Gate-based noise detection suffers on low-quality PI datasets.
- Theoretical analysis is limited to a simplified linear regression setting.
Experimental setup
Datasets: CIFAR-10H CIFAR-10N CIFAR-100N ImageNet-PI
Benchmarks: CIFAR-10H CIFAR-10N CIFAR-100N ImageNet-PI
Metrics: test accuracy AUC for wrong-label detection
Baselines: Cross-entropy ELR HET SOP TRAM TRAM++ AFM
Builds on: TRAM TRAM++ AFM ELR mixture of experts
Closest papers by meaning
- When does Privileged information Explain Away Label Noise? (ICML 2023 · similarity 0.91)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.86)
- Two Wrongs Don’t Make a Right: Combating Confirmation Bias in Learning with Label Noise (AAAI 2023 · similarity 0.86)
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.85)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.85)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.84)
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.