Understanding the trustworthiness of a prediction yielded by a classifier is
critical for the safe and effective use of AI models. Prior efforts have been
proven to be reliable on small-scale datasets. In this work, we study the
problem of predicting trustworthiness on real-world large-scale
本文提出了一种基于 Correctness Ranking Loss 的深度神经网络训练方法,可以对类别概率进行显式的正序排名,提高置信度预测。该方法易于实现,不需要额外的计算代价,适用于现有体系结构,并且在分类基准数据集上表现良好。同时,还对置信度估计相关的任务,如超出分布检测和主动学习,具有相同的可靠性。