Research map / Label-efficient and robust learning / Domain adaptation and generalization
Confidence Calibration for Domain Generalization Under Covariate Shift
Yunye Gong, Xiao Lin, Yi Yao, Thomas G. Dietterich, Ajay Divakaran, Melinda Gervasio
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 |
|---|---|---|---|
| calibrate with temperatures from source domain only instead of leaving uncalibrated (ablation) | temperature scaling · Office-Home | ECE (%): 8.66 -> 10.45 | Hurt |
| add multiple calibration domains (set-level) versus source-only calibration | temperature scaling · Office-Home | ECE (%): 10.45 -> 6.80 | Improved |
| add multiple calibration domains (set-level) versus source-only calibration | temperature scaling · DomainNet | ECE (%): 16.93 -> 7.83 | Improved |
| use cluster-level nearest-neighbor temperature instead of a single set-level temperature | set-level calibration · DomainNet | ECE (%): 7.83 -> 6.44 | Improved |
| use cluster-level nearest-neighbor temperature instead of a single set-level temperature | set-level calibration · Office-Home | ECE (%): 6.80 -> 6.98 | No clear effect |
| replace NN cluster selection with linear regression of temperature (ablation) | cluster-level calibration · DomainNet | ECE (%): 6.44 -> 9.34 | Hurt |
| ensemble logits of set, NN and regression methods | calibration methods · Office-Home | ECE (%): 6.80 -> 6.53 | Improved |
| use our multi-domain calibration instead of TransCal | TransCal · Office-Home | ECE (%): 15.39 -> 6.80 | Improved |
| use our multi-domain calibration instead of WTS | WTS · Office-Home | ECE (%): 9.78 -> 6.53 | Improved |
| add domain-invariant features to the classifier before calibration (ablation) | Ours · StarCraft2 | ECE (%): 6.30 -> 4.41 | Improved |
| change number of clusters (ablation) | cluster-level calibration · Office-Home / DomainNet | ECE (%) | No clear effect |
| train classifier on multiple source domains instead of one (ablation) | classifier · Office-Home | accuracy (%): 51.55 -> 63.35 | Improved |
Closest papers by meaning
- On Calibration and Out-of-Domain Generalization (NeurIPS 2021 · similarity 0.90)
- Pseudo-Calibration: Improving Predictive Uncertainty Estimation in Unsupervised Domain Adaptation (ICML 2024 · similarity 0.89)
- Pseudo-Labeling for Unsupervised Domain Adaptation with Kernel GLMs (NeurIPS 2026 · similarity 0.87)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.87)
- Bridging Multicalibration and Out-of-distribution Generalization Beyond Covariate Shift (NeurIPS 2024 · similarity 0.86)
- LaSCal: Label-Shift Calibration without target labels (NeurIPS 2024 · similarity 0.86)
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.
