BriefGPT.xyz
Jun, 2020
使用最大平均差异进行校准可靠回归
Calibrated Reliable Regression using Maximum Mean Discrepancy
HTML
PDF
Peng Cui, Wenbo Hu, Jun Zhu
TL;DR
本文提出了使用最大平均偏差最小化核嵌入度量的校准回归方法，证明该方法在样本量足够大时其校准误差渐近收敛于零，实验结果表明该方法能够产生良好校准和尖锐的预测区间，优于相关现有技术。
Abstract
Accurate quantification of uncertainty is crucial for real-world applications of
machine learning
. However, modern deep neural networks still produce unreliable
predictive uncertainty
, often yielding over-confide
→