NeurIPS 2020

Posterior Re-calibration for Imbalanced Datasets

Meta Review

This paper addresses the prior shift between training and testing scenarios. Using the optimal Bayes classifier, they derive a factor which is optimizes on a validation set. Reviewers found the method novel, simple to understand, easy to implement and applicable to a wide variety of tasks. It can also be combined easily with other existing domain adaptation methods. On several benchmarks, authors have achieved improved on state-of-the-art results. Thus demonstrated the general applicability of this approach. Reviewers #2 and #3 provide constructive suggestions to improve the presentation and the value of this work.