Research map / Label-efficient and robust learning / Noisy labels
Correlated Input-Dependent Label Noise in Large-Scale Image Classification
Check 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 |
|---|---|---|---|
| Model full input-dependent class covariance rather than diagonal covariance (ablation) | Heteroscedastic classifier · ImageNet ILSVRC12 val, 270 epochs | Top-1 accuracy: 78.7$^\dagger$ ($\pm 0.09$) -> 79.3 ($\pm 0.10$) | Helped |
| Model full input-dependent class covariance rather than diagonal covariance (ablation) | Heteroscedastic classifier · WebVision 1.0 val | Top-1 accuracy: 76.2$^\dagger$ ($\pm 0.15$) -> 76.6 ($\pm 0.13$) | Helped |
| Model full input-dependent class covariance rather than diagonal covariance (ablation) | Heteroscedastic PE classifier · JFT test | gAP: 64.1$^\dagger$ ($\pm 0.07$) -> 64.7 ($\pm 0.06$) | Helped |
| Tune sigmoid temperature from 1.0 to 0.15 (ablation) | Heteroscedastic diagonal classifier · ImageNet-21k test | gAP: 45.9$^\dagger$ ($\pm 0.06$) -> 46.8$^\dagger$ ($\pm 0.04$) | Helped |
| Tune sigmoid temperature from 1.0 to 0.15 (ablation) | Heteroscedastic diagonal classifier · JFT test | gAP: 63.1$^\dagger$ ($\pm 0.03$) -> 64.1$^\dagger$ ($\pm 0.07$) | Helped |
| Train longer, from 90 to 270 epochs (ablation) | Heteroscedastic full classifier · ImageNet ILSVRC12 val | Top-1 accuracy: 78.5 -> 79.3 | Helped |
| Train longer, from 90 to 270 epochs (ablation) | Homoscedastic classifier · ImageNet ILSVRC12 val | Top-1 accuracy: 76.5 -> 76.7 | Helped |
| Increase Monte Carlo samples from 1 to 10000 (ablation) | Heteroscedastic classifier · ImageNet ILSVRC12, 270 epochs | Top-1 accuracy: 78.5 -> 79.3 | Helped |
| Use an equal-parameter homoscedastic model instead of the heteroscedastic full model (ablation) | ImageNet ILSVRC12 classifier · ImageNet ILSVRC12, 270 epochs | Top-1 accuracy: 79.3 -> 77.1 | Hurt |
| Replace softmax with sigmoid (ablation) | Homoscedastic classifier · ImageNet ILSVRC12, 270 epochs | Top-1 accuracy: 76.7 ($\pm 0.13$) -> 78.2 ($\pm 0.16$) | Helped |
| Replace softmax with sigmoid (ablation) | Heteroscedastic full classifier · ImageNet ILSVRC12, 270 epochs | Top-1 accuracy: 79.3 ($\pm 0.10$) -> 79.35 ($\pm 0.05$) | No clear effect |
| Train an ensemble of four heteroscedastic models instead of one homoscedastic model (ablation) | ResNet-50 classifier · ImageNet ILSVRC12 val | Top-1 accuracy: 76.1 -> 79.5 | Helped |
Closest papers by meaning
- Learning with Feature-Dependent Label Noise: A Progressive Approach (ICLR 2021 · similarity 0.91)
- Noise-Robust Learning from Multiple Unsupervised Sources of Inferred Labels (AAAI 2022 · similarity 0.90)
- An Information Fusion Approach to Learning with Instance-Dependent Label Noise (ICLR 2022 · similarity 0.89)
- Safeguarded Dynamic Label Regression for Noisy Supervision (AAAI 2019 · similarity 0.89)
- Joint Optimization Framework for Learning With Noisy Labels (CVPR 2018 · similarity 0.89)
- On the Role of Label Noise in the Feature Learning Process (ICML 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.