RevisitingtheCalibrationof ModernNeuralNetworks

These concerns are more relevant than ever, since the architecture size, amount of training data, and computing power used by state-of-the-art models continue to increase.