Research map / Label-efficient and robust learning / Recognition and long-tail
A Geometric Perspective towards Neural Calibration via Sensitivity Decomposition
Junjiao Tian, Dylan Yung, Yen-Chang Hsu, Zsolt Kira
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 |
|---|---|---|---|
| Apply GSD with validation-set grid search of β′ | ResNet18 · CIFAR10 noise level 1 | ECE: 0.102±0.001 -> 0.040±0.002 | Improved |
| Apply GSD with validation-set grid search of β′ | ResNet18 · CIFAR100 noise level 5 | ECE: 0.237±0.01 -> 0.145±0.013 | Improved |
| Apply GSD calibration | ResNet18 · CIFAR100 rotated validation set | ECE: 0.215±0.007 -> 0.097±0.003 | Improved |
| Apply GSD calibration | ResNet18 · CIFAR10 rotated validation set | NLL: 2.941±0.085 -> 2.211±0.04 | Improved |
| Disentangle the feature norm and regularize α near one | ResNet18 · CIFAR10 vs. SVHN | AUROC using norm: 90.48 -> 99.05 | Improved |
| Disentangle the feature norm and regularize α near one | ResNet18 · CIFAR100 vs. SVHN | AUROC using norm: 79.38 -> 94.46 | Improved |
| Remove input feature norm (ablation) | Vanilla ResNet18 · CIFAR10, clean | ECE: 0.025±0.001 -> 0.893±0.002 | Hurt |
| Remove classifier weight norm (ablation) | Vanilla ResNet18 · CIFAR10, clean | ECE: 0.025±0.001 -> 0.061±0.0003 | Hurt |
| Remove both classifier weight norm and input feature norm (ablation) | Vanilla ResNet18 · CIFAR10, clean | ECE: 0.025±0.001 -> 0.914±0.001 | Hurt |
| Remove regularization that keeps α near one (ablation) | GSD ResNet18 · CIFAR100 vs. SVHN | AUROC using cosine similarity: 85.68 -> 72.61 | Hurt |
Closest papers by meaning
- Learning for Transductive Threshold Calibration in Open-World Recognition (CVPR 2024 · similarity 0.84)
- Invertible Logits Transformation for Accuracy-Preserving Post-Hoc Uncertainty Calibration (NeurIPS 2026 · similarity 0.84)
- Divide et Calibra: Multiclass Local Calibration via Vector Quantization (NeurIPS 2026 · similarity 0.83)
- Decoupling Feature Extraction and Classification Layers for Calibrated Neural Networks (ICML 2024 · similarity 0.83)
- Calibration Bottleneck: Over-compressed Representations are Less Calibratable (ICML 2024 · similarity 0.83)
- Provable Uncertainty Decomposition via Higher-Order Calibration (ICLR 2025 · similarity 0.83)
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.
