Research map / Architectures and efficiency / Quantization
Precision-Induced Miscalibration: Understanding and Correcting Confidence Distortion in Quantized Neural Networks
Jiawei Gu, Fengyuan Nie, Hao Tang, Yanpeng Sun
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 |
|---|---|---|---|
| add PACS sample-adaptive temperature at inference | ResNet-50 FP16 · ImageNet-1K val | ECE (%): 5.82 -> 1.92 | Improved |
| use adaptive PACS temperature instead of fixed global T=1.5 (ablation) | PACS · ImageNet-1K val ResNet-50 | ECE (%): 3.45 -> 1.92 | Improved |
| use true precision risk score instead of random risk (ablation) | PACS · ImageNet-1K val ResNet-50 | ECE (%): 4.87 -> 1.92 | Improved |
| use L-inf norm for logit scale instead of L2 (ablation) | PACS · ImageNet-1K val ResNet-50 | ECE (%): 2.34 -> 1.92 | Improved |
| add stop-gradient on T(x) (ablation) | PACS (training) · ImageNet ResNet-50 | ECE (%) / stability: 2.34 -> 1.92; 3 of 10 runs NaN without | Improved |
| add PACS on top of Label Smoothing | Label Smoothing · ImageNet-1K val ResNet-50 | ECE (%): 3.21 -> 1.54 | Improved |
| add PACS on top of Logit Normalization | LN · ImageNet-1K val ResNet-50 | ECE (%): 2.05 -> 1.75 | Improved |
| apply Unit Scaling | Standard FP16 · ImageNet-1K val ResNet-50 | ECE (%): 5.82 -> 5.65 | No clear effect |
| apply PACS in INT8 post-training quantization | ResNet-50 · ImageNet-1K val | ECE (%): 5.23 -> 4.56 (vs LN) | Improved |
| apply PACS on FP32 precision | ResNet-50 · ImageNet-1K val | ECE (%): 3.18 vs 3.18 (LN) | No clear effect |
| apply PACS to LLaMA-7B BF16 | LLaMA-7B · C4 | ECE: 12.3% -> 7.8% | Improved |
| run trunk in FP16 with FP32 softmax vs fully FP32 | ResNet-50 · ImageNet-1K val | ECE degradation (points): +1.68 (trunk), +0.57 (softmax), +2.61 (both) | Hurt |
Closest papers by meaning
- Sample Margin-Aware Recalibration of Temperature Scaling (ICML 2026 · similarity 0.87)
- Scaling Laws for Precision in High-Dimensional Linear Regression (ICML 2026 · similarity 0.87)
- Quantile‑Adaptive Temperature Scaling for Confidence Calibration (ECCV 2026 · similarity 0.87)
- On Calibration of Modern Neural Networks (ICML 2017 · similarity 0.87)
- Better Uncertainty Calibration via Proper Scores for Classification and Beyond (NeurIPS 2022 · similarity 0.87)
- A Benchmark Study on Calibration (ICLR 2024 · 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.
