Research map / Label-efficient and robust learning / Noisy labels
When does Privileged information Explain Away Label Noise?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| augment TRAM with a unique random vector for each example (TRAM++) (ablation) | TRAM · CIFAR-10H (worst) | accuracy: 64.9\scriptstyle{\pm 0.8} -> 66.8\scriptstyle{\pm 0.3} | Helped |
| augment TRAM with a unique random vector for each example (TRAM++) (ablation) | TRAM · CIFAR-10N (worst) | accuracy: 80.5\scriptstyle{\pm 0.5} -> 83.9\scriptstyle{\pm 0.2} | Helped |
| augment TRAM with a unique random vector for each example (TRAM++) (ablation) | TRAM · CIFAR-100N | accuracy: 59.7\scriptstyle{\pm 0.3} -> 61.1\scriptstyle{\pm 0.2} | Helped |
| augment TRAM with a unique random vector for each example (TRAM++) (ablation) | TRAM · ImageNet-PI (high-noise) | accuracy: 53.3 \scriptstyle{\pm 0.5} -> 53.9 \scriptstyle{\pm 0.4} | Helped |
| combine TRAM++ with SOP fine-tuning | TRAM++ · CIFAR-10H (worst) | accuracy: 59.2 \scriptstyle{\pm 0.8} -> \bm{70.9\scriptstyle{\pm 0.5}} | Helped |
| combine TRAM++ with SOP fine-tuning | TRAM++ · CIFAR-10N (worst) | accuracy: 87.9 \scriptstyle{\pm 0.2} -> \bm{88.5\scriptstyle{\pm 0.3}} | Helped |
| combine TRAM++ with SOP fine-tuning | TRAM++ · CIFAR-100N | accuracy: 65.3 \scriptstyle{\pm 0.3} -> \bm{66.1\scriptstyle{\pm 0.2}} | Helped |
| apply heteroscedastic output layers to TRAM++ | TRAM++ · ImageNet-PI (high-noise) | accuracy: 51.5 \scriptstyle{\pm 0.6} -> \bm{55.8 \scriptstyle{\pm 0.3}} | Helped |
| use fully target-predictive PI (noisy labels as PI) instead of original PI (ablation) | TRAM and Approximate FM · CIFAR-10H (worst) | accuracy: 64.9\scriptstyle{\pm 0.8} -> 38.3\scriptstyle{\pm 0.2}; 64.0\scriptstyle{\pm 0.6} -> 29.5\scriptstyle{\pm 0.5} | Hurt |
| use fully target-predictive PI (noisy labels as PI) instead of original PI (ablation) | TRAM and Approximate FM · CIFAR-10N (worst) | accuracy: 80.5\scriptstyle{\pm 0.5} -> 48.9\scriptstyle{\pm 0.2}; 82.0\scriptstyle{\pm 0.3} -> 22.6\scriptstyle{\pm 0.2} | Hurt |
| use fully target-predictive PI (noisy labels as PI) instead of original PI (ablation) | TRAM and Approximate FM · CIFAR-100N | accuracy: 59.7\scriptstyle{\pm 0.3} -> 34.9\scriptstyle{\pm 0.2}; 60.0\scriptstyle{\pm 0.2} -> 20.1\scriptstyle{\pm 0.3} | Hurt |
| use fully target-predictive PI (noisy labels as PI) instead of original PI (ablation) | TRAM and Approximate FM · ImageNet-PI (high-noise) | accuracy: 53.3\scriptstyle{\pm 0.5} -> 41.0\scriptstyle{\pm 0.7}; 55.6\scriptstyle{\pm 0.3} -> 0.8\scriptstyle{\pm 0.2} | Hurt |
Closest papers by meaning
- Pi-DUAL: Using privileged information to distinguish clean from noisy labels (ICML 2024 · similarity 0.91)
- Transfer and Marginalize: Explaining Away Label Noise with Privileged Information (ICML 2022 · similarity 0.89)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.87)
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.86)
- Understanding Instance-Level Label Noise: Disparate Impacts and Treatments (ICML 2021 · similarity 0.86)
- Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks (ICLR 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.