BriefGPT.xyz
Oct, 2022
标定泛化差距
The Calibration Generalization Gap
HTML
PDF
Annabelle Carrell, Neil Mallinar, James Lucas, Preetum Nakkiran
TL;DR
通过将校准误差分解为训练集的校准误差和校准泛化间隙,我们理论证明了深度神经网络在训练集上通常是校准的,校准泛化间隙受到标准泛化间隙的限制,因此具有小的泛化间隙的模型是校准的。
Abstract
calibration
is a fundamental property of a good
predictive model
: it requires that the model predicts correctly in proportion to its confidence. Modern
→