Research map / Label-efficient and robust learning / Recognition and long-tail
Target-Agnostic Calibration under Distribution Shift with Frequency-Aware Gradient Rectification
Yilin Zhang, Cai Xu, You Wu, Ziyu Guan, Wei Zhao
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 low-pass filtering without gradient rectification (Filter Only) (ablation) | FGR · CIFAR-10 ID | ECE (%): 0.65 -> 11.54 | Hurt |
| add low-pass filtering to rectification (vs Rect. Only) (ablation) | FGR · CIFAR-10-C | ECE (%): 18.60 -> 6.78 | Improved |
| add gradient rectification to filtering (vs Filter Only) (ablation) | FGR · CIFAR-10 ID | ECE (%): 11.54 -> 0.65 | Improved |
| replace one-sided projection with weighted loss sum (ablation) | FGR · CIFAR-100-C | ECE (%): 10.50 -> 11.80 | Hurt |
| use one-sided rectification instead of PCGrad (ablation) | FGR · Camelyon17 test | ECE (%): 3.38 -> 1.14 | Improved |
| use one-sided rectification instead of CAGrad (ablation) | FGR · Camelyon17 test | ECE (%): 1.28 -> 1.14 | Improved |
| use PCGrad instead of FGR rectification (ablation) | FGR · iWildCam test | accuracy (%): 75.22 -> 77.21 | Improved |
| start FGR too early in training (ablation) | FGR · CIFAR-10/100 | ID accuracy | Hurt |
| increase filtering ratio rho (ablation) | FGR · CIFAR-10/100 and iWildCam | shifted calibration vs clean performance | Mixed |
| vary DCT strength lambda within {15,18,25} (ablation) | FGR · CIFAR-10/100 and iWildCam | calibration | No clear effect |
| two-stage fine-tuning (frozen backbone) instead of from-scratch FGR (ablation) | FGR · CIFAR-10 ID (ResNet-50) | ECE (%): 0.65 -> 0.58 | Improved |
| apply FGR to transformer backbones instead of CE | Swin-Tiny · Camelyon17-OOD | ECE (%): 15.96 -> 8.21 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- AugMix — AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (ICLR 2020) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Uncertainty Weighted Gradients for Model Calibration (CVPR 2025 · similarity 0.87)
- Calibration Bottleneck: Over-compressed Representations are Less Calibratable (ICML 2024 · similarity 0.86)
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.85)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.85)
- Divide et Calibra: Multiclass Local Calibration via Vector Quantization (NeurIPS 2026 · similarity 0.85)
- Towards Calibrated Robust Fine-Tuning of Vision-Language Models (NeurIPS 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.
