BriefGPT.xyz
Oct, 2023
如果没有欠拟合,就没有冷后验效应
If there is no underfitting, there is no Cold Posterior Effect
HTML
PDF
Yijie Zhang, Yi-Shan Wu, Luis A. Ortega, Andrés R. Masegosa
TL;DR
贝叶斯深度学习中的冷后验效应表明,在温度$ T < 1 $的后验中,预测效果可能比贝叶斯后验($ T = 1 $)要好。本研究更深入地阐述了冷后验效应,揭示只有当贝叶斯后验出现欠拟合情况时,才会出现冷后验效应。事实上,理论上证明了如果没有欠拟合,就不会有冷后验效应。
Abstract
The
cold posterior effect
(CPE) (Wenzel et al., 2020) in
bayesian deep learning
shows that, for posteriors with a
temperature
$T<1$, the r
→