Neighborhood Smoothing for Calibration

Idan Horowitz, Avigdor Gal

Modern neural networks are often miscalibrated, with a tendency to overconfidence. Existing train-time calibration methods largely modify task losses or calibration penalties, leav…