Improving DNN Reliability via Refinement and Calibration

A new arXiv preprint (2605.23249) addresses the unreliability of confidence estimates in deep neural networks (DNNs), which can undermine user trust. While calibration—aligning predicted confidence with actual correctness—has been a focus, the authors note that post-hoc calibration methods often reduce refinement (sharpness), i.e., the ability to assign distinct confidence scores to correct versus incorrect predictions. They propose a method to improve both calibration and refinement, aiming for more reliable DNNs without sacrificing predictive accuracy.