Adaptive Class-Balanced Loss Based on Re-Weighting

Abstract

As real-world data grows fast, the problem of data imbalance has become more prominent. Thus the long-tail problem in deep learning has received lots of attention recently. One of the solutions is to apply a class rebalancing strategy, such as directly using the inverse of the class sample size for reweighting. In past studies, the setting of weights only relates to the number of class samples. Only relying on the information of the number of class samples to determine the size of the weight is very crude in the sensitive method of re-weighting. In this paper, we implement adaptive re-weighting for three essential attributes of the dataset considering several factors: the number of classes, the number of samples, and the degree of class imbalance. We conducted experiments on the commonly used sample imbalance problem solution and proposed a new sample reweighting method. Specifically, a novel re-weighting idea is proposed to optimize Class-Balanced Loss Based on an Effective Number of Samples. Experiments show that the method is superior in re-weighting imbalanced datasets on deep neural networks. We hope our work will stimulate a rethinking of the number-of-samples-based convention in re-weighting.