Research map / Architectures and efficiency / RNNs and normalization
Revisiting the Calibration of Modern Neural Networks
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 |
|---|---|---|---|
| increase pretraining dataset size 300-fold (ImageNet to JFT-300M) (ablation) | BiT models · ImageNet, temperature-scaled | ECE: 0.0185 -> 0.0185 (R50x1); 0.0272 -> 0.0236 (R101x3) | No clear effect |
| increase pretraining dataset size (ablation) | BiT models · ImageNet | classification error | Helped |
| longer pretraining duration (ablation) | BiT-R101x3 · ImageNet | ECE: 0.019 -> 0.028 | Hurt |
| longer pretraining duration (ablation) | BiT-R101x3 · ImageNet | classification error: 0.232 -> 0.176 | Helped |
| increase model size within family | ViT · ImageNet | ECE: 0.01 -> 0.017 | Hurt |
| increase model size within family | ViT · ImageNet | classification error: 0.194 -> 0.118 | Helped |
| increase model size under distribution shift | image classifiers · ImageNet-C high severity | ECE | Helped |
| apply temperature scaling | image classifiers · ImageNet | ECE | Helped |
| add label smoothing during training (ablation) | EfficientNet-B4 · ImageNet-C | ECE after temperature scaling | Hurt |
| add label smoothing during training (ablation) | EfficientNet-B4 · ImageNet-C | ECE before temperature scaling | Mixed |
| choose higher-accuracy model over better-calibrated model in selective prediction | BiT-R152x4 vs BiT-R50x1 · selective prediction cost | expected cost: classification error better by 0.08 vs ECE better by 0.009 | Helped |
Closest papers by meaning
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.91)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.90)
- Calibration Bottleneck: Over-compressed Representations are Less Calibratable (ICML 2024 · similarity 0.89)
- Uncertainty Weighted Gradients for Model Calibration (CVPR 2025 · similarity 0.88)
- Improving GNN Calibration with Discriminative Ability: Insights and Strategies (AAAI 2024 · similarity 0.88)
- Decoupling Feature Extraction and Classification Layers for Calibrated Neural Networks (ICML 2024 · similarity 0.87)
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.