Research map / Label-efficient and robust learning / Recognition and long-tail
From Individual Calibration to Reliable Classifiers: ALD Parameterization with mPAIC Guarantees
Deming Sheng, Ricardo Henao
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 |
|---|---|---|---|
| replace Softmax with the asymmetric Laplace distribution (ALD) classifier (ablation) | distribution-based classifier · 9 datasets (Figure 2a) | overall average rank: 4.21 -> 3.00 | Improved |
| replace Softmax with ALD classifier (ablation) | distribution-based classifier · 9 datasets, ECE rank (Figure 2b) | ECE average rank: 4.58 -> 3.07 | Improved |
| add ICALD individualized calibration objective (pre-calibration) to the ALD classifier | HALD · 9 datasets, 45 settings | overall average rank: 6.10 -> 4.65 | Improved |
| add ICALD calibration as post-calibration adapter instead of temperature scaling | HALD · 9 datasets, 45 settings | overall average rank: HALD+TS 7.23 -> HPost-ICALD 5.03 | Improved |
| add ICALD pre-calibration to ALD vs MMD distribution matching | HALD · 9 datasets, KCE rank | KCE average rank: HALD+MMD 3.49 vs HPre-ICALD 4.24 | Hurt |
| add post-hoc temperature scaling to Softmax | HSoftmax · 9 datasets, 45 settings | overall average rank: 7.11 -> 7.74 | Hurt |
| add post-hoc histogram binning to Softmax | HSoftmax · 9 datasets, 45 settings | overall average rank: 7.11 -> 8.72 | Hurt |
| add MMD pre-calibration to Softmax | HSoftmax · 9 datasets, 45 settings | overall average rank: 7.11 -> 6.23 | Improved |
| allow calibration gradients to update the backbone (Full Grad) instead of detaching them (ablation) | HALD · Fisher Ratio on BREAST-CANCER, SST-2 | Fisher Ratio | Hurt |
| add gradient detachment (DDSO) to calibration stream (ablation) | HICALD · Fisher Ratio across datasets | Fisher Ratio | Improved |
| add gradient detachment (DDSO) to calibration stream (ablation) | HICALD · training cost | relative train+inference time: 1x -> 0.554x | Improved |
| use ALD instead of Softmax | classifier head · training time relative to Softmax | relative wall-clock time: 1.020x | No clear effect |
Closest papers by meaning
- Divide et Calibra: Multiclass Local Calibration via Vector Quantization (NeurIPS 2026 · similarity 0.87)
- Balancing Two Classifiers via A Simplex ETF Structure for Model Calibration (CVPR 2025 · similarity 0.87)
- Improving Multi-Class Calibration through Normalization-Aware Isotonic Techniques (ICML 2025 · similarity 0.86)
- Confidence Calibration of Classifiers with Many Classes (NeurIPS 2024 · similarity 0.86)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.86)
- The Minimax Rate of Second-Order Calibration (NeurIPS 2026 · 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.
