Research map / Label-efficient and robust learning / Recognition and long-tail
Rethinking Calibration of Deep Neural Networks: Do Not Be Afraid of Overconfidence
Deng-Bao Wang, Lei Feng, Min-Ling Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add regularization (label smoothing, L1 norm, focal loss) in training instead of plain cross-entropy | ResNet-32 · CIFAR-10 | ECE without post-hoc calibration: 6.43 -> 2.72 (label smoothing) | Improved |
| add regularization (label smoothing) in training then apply temperature scaling | ResNet-32 · CIFAR-10 | ECE with TS: 0.95 -> 2.54 | Hurt |
| add focal loss in training then apply temperature scaling | ResNet-32 · CIFAR-100 | ECE with TS: 1.35 -> 2.14 | Hurt |
| add L1 norm regularization then apply temperature scaling | 8-layer 1D-CNN · 20 Newsgroups | ECE with TS: 3.11 -> 2.71 | Improved |
| add label smoothing or focal loss then apply histogram binning | ResNet-32 · CIFAR-10 | ECE with HB: 0.74 -> 0.94 (LS); 0.74 -> 1.65 (FL) | Hurt |
| add focal loss with strong coefficient | ResNet-32 · CIFAR-10 | Accuracy: 90.46 -> 87.84 | Hurt |
| replace cross-entropy with inverse focal loss (weights confident samples more) plus temperature scaling | ResNet-32 · CIFAR-10 and CIFAR-100 | ECE after TS | Improved |
| replace cross-entropy with inverse focal loss, histogram binning | ResNet-32 · SVHN, CIFAR-10, CIFAR-100 | ECE after HB | Mixed |
| replace cross-entropy with inverse focal loss, no post-hoc calibration | ResNet-32 · most datasets | ECE without post-hoc | Hurt |
| replace cross-entropy with inverse focal loss, large coefficient | 8-layer 1D-CNN · 20 Newsgroups | Accuracy | Hurt |
Closest papers by meaning
- Better Uncertainty Calibration via Proper Scores for Classification and Beyond (NeurIPS 2022 · similarity 0.91)
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.90)
- Calibration Bottleneck: Over-compressed Representations are Less Calibratable (ICML 2024 · similarity 0.90)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.89)
- Uncertainty Weighted Gradients for Model Calibration (CVPR 2025 · 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.
