New method enables conformal prediction transfer across domains

Researchers have developed a novel method called Transported Conformal Calibration (TCC) to address the challenge of conformal prediction when labeled calibration data is only available in a source domain, but predictions are needed in a target domain. TCC transfers labeled source calibration to the target space using unlabeled paired observations and then corrects for any residual mismatch using only unlabeled target inputs. The proposed methods, TCC-KS and weighted-TCC, offer finite-sample coverage guarantees that adapt to observable mismatch, demonstrating reliable transfer across datasets like CIFAR-100-C and Tiny-ImageNet-C without requiring labeled target calibration data. AI

IMPACT Enables more robust uncertainty quantification in machine learning models when domain shift occurs.

RANK_REASON Academic paper detailing a new machine learning method. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• CIFAR-100-C
• Conformal Calibration Transfer
• SEN12MS
• TCC-KS
• Tiny-ImageNet-C
• Transported Conformal Calibration
• weighted-TCC

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →