Research map / Label-efficient and robust learning / Recognition and long-tail
Residual Calibration via Local Feature-Space Refinement
Xihao Wang, Chengyao Yu, Ruixing Ming
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a post-hoc calibration method that corrects spatially varying miscalibration by refining predictions in local regions of the learned feature space rather than applying a global score-based transformation.
Problem
Deep neural networks give unreliable confidence estimates; existing post-hoc calibration methods are score-based and improve marginal calibration but miss spatially varying miscalibration in local feature-space regions.
Approach
Shifts post-hoc calibration from score-space transformations to local refinement in feature space, estimating calibration adjustments per local region (the abstract is truncated, so the exact estimator is not stated).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CFAF-LAPE instead of late fusion for multi-view calibration | multi-view calibration · Synthetic XOR views | NLL: 0.507 -> 0.413 | Improved |
| Use CFAF-LAPE instead of late fusion | multi-view calibration · Multi-view Digits | NLL: 0.112 -> 0.092 | Improved |
| Use CFAF-LAPE instead of late fusion | multi-view calibration · HAM10000 melanoma vs nevus | ECE: 0.023 -> 0.014 | Improved |
| Use CFAF-LAPE instead of late fusion | multi-view calibration · Multi-view Digits | accuracy: 0.960 -> 0.964 | No clear effect |
| Add LAPE residual refinement to TS, VS, FC and SAG anchors | post-hoc calibrators · Obesity Risk, CIFAR-100, BreastMNIST, MHIST | NLL and ECE | Mixed |
| Use local label frequency from nearest calibration neighbors instead of softmax | probability estimation · 2D simulation, softmax scores in [0.35,0.65] | MAE to oracle probability: 0.156 -> 0.126 | Improved |
| Use local frequency alone as a replacement for softmax over all test points (ablation) | probability estimation · 100-dim simulation, entire test set | MAE to oracle probability: 0.0941 -> 0.1058 | Hurt |
| Use local frequency in the predefined boundary region of 100-dim simulation | probability estimation · 100-dim simulation, boundary region | MAE to oracle probability: 0.1067 -> 0.0969 | Improved |
| Add more calibration data (ablation) | VS+LAPE · calibration-ratio sweep | NLL | Improved |
| Use V-fold cross-fitting for anchor scores instead of direct reuse or data splitting (ablation) | LAPE · flexible anchors, limited calibration data | calibration quality | Improved |
Closest papers by meaning
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.86)
- Divide et Calibra: Multiclass Local Calibration via Vector Quantization (NeurIPS 2026 · similarity 0.86)
- Proximity-Informed Calibration for Deep Neural Networks (NeurIPS 2023 · similarity 0.86)
- Calibration Bottleneck: Over-compressed Representations are Less Calibratable (ICML 2024 · similarity 0.85)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.84)
- Uncertainty Weighted Gradients for Model Calibration (CVPR 2025 · similarity 0.84)
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.
