Research map / Label-efficient and robust learning / Recognition and long-tail
Intra Order-preserving Functions for Calibration of Multi-Class Neural Networks
Amir Rahimi, Amirreza Shaban, Ching-An Cheng, Richard Hartley, Byron Boots
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 |
|---|---|---|---|
| learn calibration function within intra order-preserving family (Diag) instead of temperature scaling | post-hoc calibration · 14 dataset-model pairs | average relative ECE: 0.42 -> 0.27 | Improved |
| use order-invariant order-preserving function (OI) instead of temperature scaling | post-hoc calibration · 14 dataset-model pairs | average relative ECE: 0.42 -> 0.33 | Improved |
| add order-preserving constraint (OP) vs unconstrained MLP calibrator (ablation) | MLP calibration function · 14 dataset-model pairs | average relative ECE: 0.66 -> 0.41 | Improved |
| unconstrained MLP calibrator hurts accuracy (ablation) | MLP calibration function · ImageNet ResNet-152 | accuracy change: -33.4% | Hurt |
| add order-invariant sharing (OI) to general order-preserving (OP) (ablation) | intra order-preserving calibrator · 14 dataset-model pairs | average relative ECE: 0.41 -> 0.33 | Improved |
| add diagonal parameter sharing (Diag) to OI (ablation) | intra order-preserving calibrator · 14 dataset-model pairs | average relative ECE: 0.33 -> 0.27 | Improved |
| Diag vs Dirichlet calibration on classwise-ECE | post-hoc calibration · average relative Classwise-ECE | relative Classwise-ECE: 0.704 -> 0.729 | Hurt |
| Diag vs temperature scaling on NLL | post-hoc calibration · average relative NLL | relative NLL: 0.766 -> 0.749 | Improved |
Closest papers by meaning
- Meta-Cal: Well-controlled Post-hoc Calibration by Ranking (ICML 2021 · similarity 0.88)
- Post-Hoc Uncertainty Calibration for Domain Drift Scenarios (CVPR 2021 · similarity 0.87)
- Rethinking Calibration of Deep Neural Networks: Do Not Be Afraid of Overconfidence (NeurIPS 2021 · similarity 0.87)
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.87)
- Calibration Bottleneck: Over-compressed Representations are Less Calibratable (ICML 2024 · similarity 0.86)
- Uncertainty Weighted Gradients for Model Calibration (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.
