Research map / Architectures and efficiency / RNNs and normalization
On 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 model depth or width (ablation) | ResNet on CIFAR-100 · CIFAR-100 | ECE | Hurt |
| Add Batch Normalization (ablation) | 6-layer ConvNet · Image classification | ECE | Hurt |
| Reduce weight decay (ablation) | 110-layer ResNet · CIFAR-100 | ECE | Hurt |
| Increase weight decay (ablation) | 110-layer ResNet · CIFAR-100 | ECE | Helped |
| Apply temperature scaling | Neural network classifiers · Vision classification tasks | ECE | Helped |
| Apply temperature scaling | Neural network classifiers · NLP classification tasks | ECE | Helped |
| Apply temperature scaling | Neural network classifiers · Reuters | ECE: ECE ≤ 1% | No clear effect |
| Use matrix scaling instead of temperature scaling | Neural network classifiers · Birds, Cars, and CIFAR-100 | ECE | Hurt |
| Use matrix scaling | Neural network classifier · ImageNet | convergence | Hurt |
| Use binning methods rather than temperature scaling | Neural network classifiers · Most evaluated datasets | ECE | Hurt |
| Apply histogram binning, isotonic regression, or BBQ | Neural network classifiers · Most evaluated datasets | accuracy | Hurt |
| Continue training into the region where NLL overfits | Neural network on CIFAR-100 · CIFAR-100 | test error: 29% to 27% | Helped |
Closest papers by meaning
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.91)
- Confidence Calibration of Classifiers with Many Classes (NeurIPS 2024 · similarity 0.89)
- On the Limitations of Temperature Scaling for Distributions with Overlaps (ICLR 2024 · similarity 0.89)
- On Temperature Scaling and Conformal Prediction of Deep Classifiers (ICML 2025 · similarity 0.88)
- Uncertainty Weighted Gradients for Model Calibration (CVPR 2025 · similarity 0.88)
- Parametric ρ-Norm Scaling Calibration (AAAI 2025 · 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.