Research map / Learning theory and algorithms / Prediction and decision losses
Sample Margin-Aware Recalibration of Temperature Scaling
Haolan Guo, Linwei Tao, Haoyang Luo, Minjing Dong, Chang Xu
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 |
|---|---|---|---|
| train the margin-to-temperature map with Charbonnier-SoftECE instead of NLL (ablation) | SMART · ImageNet-1K ViT-B/16 | ECE (%): 3.62 -> 0.48 | Improved |
| train with Charbonnier-SoftECE instead of NLL (ablation) | SMART · ImageNet-1K ResNet-50 | ECE (%): 0.93 -> 0.52 | Improved |
| add Charbonnier smoothing to SoftECE (ablation) | SMART · ImageNet-1K | ECE (%): ResNet-50 0.65 -> 0.52; ViT-B/16 0.89 -> 0.48 | Improved |
| use logit margin as input instead of entropy, confidence, all logits, max logit or max-mean (ablation) | SMART · ImageNet-1K | ECE (%): ResNet-50 0.85-0.97 -> 0.52 | Improved |
| NLL training of the temperature network for longer (80 epochs) (ablation) | SMART with NLL objective · ImageNet ViT-B/32 | ECE: NLL down 15%, ECE up 8% relative to epoch 30 | Hurt |
| replace global temperature (TS) with SMART per-sample margin-based temperature | Temperature Scaling · CIFAR-100 ResNet-50 | ECE (%): 5.61 -> 1.37 | Improved |
| apply SMART post-hoc on top of models trained with calibration losses | NLL, Brier, MMCE, LS-0.05, FLSD-53, FL-3 trained models · CIFAR-10 / CIFAR-100 | ECE (%): CIFAR-100 ResNet-50 MMCE 15.31 -> 0.86 | Improved |
| use SMART (post-hoc) instead of training-time UWG/DFL/MDCA | Vanilla classifiers · ImageNet / Sketch | ECE (%): ImageNet ViT-B/16 UWG 6.29, DFL 7.62, MDCA 10.79 vs SMART 0.48 | Improved |
| use Spline calibration under distribution shift on CNNs | Spline · ImageNet-Sketch ResNet-50 | ECE (%): SMART 0.92 vs Spline 9.76 | Hurt |
| increase validation set size from 2 to 512 samples (ablation) | SMART · ImageNet ViT-B/32 | ECE (%) | Improved |
| add more validation data to TS (ablation) | Temperature Scaling · ImageNet ViT-B/32 | ECE (%) | No clear effect |
| use SMART instead of PTS for runtime | PTS · ImageNet ResNet-50 runtime | runtime (s): 1050.44 -> 23.03 | Improved |
Closest papers by meaning
- On the Limitations of Temperature Scaling for Distributions with Overlaps (ICLR 2024 · similarity 0.88)
- Consistency-Guided Temperature Scaling Using Style and Content Information for Out-of-Domain Calibration (AAAI 2024 · similarity 0.87)
- On Temperature Scaling and Conformal Prediction of Deep Classifiers (ICML 2025 · similarity 0.87)
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.86)
- HoTS: Homophily-Aware Temperature Scaling for Graph Neural Network Calibration (NeurIPS 2026 · similarity 0.85)
- T-CIL: Temperature Scaling using Adversarial Perturbation for Calibration in Class-Incremental Learning (CVPR 2025 · similarity 0.85)
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.
