Enhancing Confidence Calibration in Long-Tailed Recognition

Abstract

Deep neural networks often struggle with heavily class-imbalanced training datasets. Recently, two-stage methods have been developed to separate representation learning from classifier learning, aiming to enhance performance. However, the crucial issue of miscalibration remains. To tackle this, we introduce novel methods to improve both calibration and performance in such scenarios. Recognizing that predicted probability distributions of classes are closely tied to the number of class instances, we propose label-aware smoothing with balanced softmax. This strategy tackles the issue of differing levels of over-confidence among different categories, thereby improving the learning process of classifiers. Furthermore, to counteract potential bias in the dataset between the two stages caused by different sampling techniques, we incorporate shifted batch normalization into the decoupling framework. The methods we suggest have set fresh standards on numerous prevalent longtailed recognition datasets such as CIFAR10-LT and CIFAR100-LT.