[Research] Recursive Lossy Label-Invariant Calibration (ReCal)

We propose a post-hoc calibration to address the mis-calibration issue of modern neural networks. We define a lossy label-invariant transformation and utilize the proporties of the transformation to calibrate neural network classifiers. We experiementally show that our approach works well and scales using several image datasets/models. This research result is published in AISTATS 2021, and the code is publicly released. Paper Github