Posterior Re-calibration for Imbalanced Datasets

Neural Networks can perform poorly when the training label distribution is heavily imbalanced, as well as when the testing data differs from the training distribution. In order to deal with shift in the testing label distribution, which imbalance causes, we motivate the problem from the perspective of an optimal Bayes classifier and derive a post-training prior rebalancing technique that can be solved through a KL-divergence based optimization. This method allows a flexible post-training hyper-parameter to be efficiently tuned on a validation set and effectively modify the classifier margin to deal with this imbalance. We further combine this method with existing likelihood shift methods, re-interpreting them from the same Bayesian perspective, and demonstrating that our method can deal with both problems in a unified way. The resulting algorithm can be conveniently used on probabilistic classification problems agnostic to underlying architectures. Our results on six different datasets and five different architectures show state of art accuracy, including on large-scale imbalanced datasets such as iNaturalist for classification and Synthia for semantic segmentation. Please see https://github.com/GT-RIPL/UNO-IC.git for implementation. Related Work In this paper, we focus on prior (label) distribution shift resulting from various degrees of imbalance in the training label distributions, and testing on a different distribution or emphasizing a different distribution in the testing metrics (e.g. class-averaged accuracy). We therefore summarize methods for dealing with such imbalance. Imbalance -Data Level Methods The simplest methods for dealing with label imbalance during training are random under-sampling (RUS) which discards samples from the majority classes and random over-sampling (ROS) which re-samples from the minority classes [3] . While ROS is infeasible when the data imbalance is extreme, RUS tends to overfit the minority classes [4] . Synthetic