Get Started
Home
Topics
Search
Library
Research questionHow can multiclass post-hoc calibration improve confidence without changing top-1 predictions?Post-hoc calibration can alter the class selected by the highest probability even when its goal is only to correct confidence. Accuracy does not reveal how frequently those top-1 decisions changed.
AI
Computer Vision
Machine Learning
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc CalibrationThe evidence concerns image classification on CIFAR-10, CIFAR-100, and ImageNet-1K, including distribution shifts and different calibration-set sizes. The proposed output repair uses the original and calibrated probability vectors; results are reported with TPCR, ECE, NLL, and Brier scores.research paper · Sep 2, 2026
Related questions
How can teams estimate whether multiclass accuracy gains reduce optimization regret before building a model?How can federated learning protect confidence calibration from attacks that preserve accuracy?How can extracted rules from tree ensembles retain calibrated probabilities while simplifying predictions?How can post-hoc OOD detection capture hierarchical features without sacrificing scale invariance or modifying the model?