May, 2022
区分最大化损失:通过替换损失和校准来有效地提高越界检测和不确定性估计
Distinction Maximization Loss: Efficiently Improving Out-of-Distribution Detection and Uncertainty Estimation by Replacing the Loss and Calibrating
David Macêdo, Cleber Zanchettin, Teresa Ludermir
TL;DR该论文提出了使用 DisMax 损失训练确定性神经网络的方法,提高了分类准确度、样本在分布内和外的识别能力和不确定性估计,同时还保持了神经网络的推理效率。