||
Minimizes augmented error, where the added regularizer effectively limits model complexity.
1. Model Selection Problem
机器学习需要太多的选择,如迭代次数,学习率的大小,损失函数,正则化参数等等。
我们的终极目标是选择能够使$E_{out}$最小的分类器,但实际上不可行。
那么,我们可以转而选择使得在测试集上次错误最小的分类器。
Comparison between $E_{in}$ and $E_{test}$
2. Validation
Model Selection by Best $E_{val}$
3. Leave-One-Out Cross Validation
Theoretical Guarantee of Leave-One-Out Estimate
4. V-Fold Cross Validation
Archiver|手机版|科学网 ( 京ICP备07017567号-12 )
GMT+8, 2024-11-24 03:42
Powered by ScienceNet.cn
Copyright © 2007- 中国科学报社