BriefGPT.xyz
Mar, 2019
弱特征的双下降模型
Two models of double descent for weak features
HTML
PDF
Mikhail Belkin, Daniel Hsu, Ji Xu
TL;DR
本文提出了“双下降”风险曲线,用于定性描述可变参数机器学习模型的样本外预测准确性。通过对最小二乘 / 最小规范预测器的两个简单数据模型进行精确的数学分析,显示出风险在特征数$p$接近样本数$n$时达到峰值,但随着$p$超过$n$而逐渐降低至最小值。而这种行为则和需要提前确定最优特征序列的“先知”模型有所不同。
Abstract
The "
double descent
"
risk curve
was recently proposed to qualitatively describe the
out-of-sample prediction
accuracy of variably-paramete
→