Regression
step1:model
y=b+w*x cp
step2:goodness of function
loss function:how bad the function is
step3:best function
令Lf最小
GRADIENT DESCENT：协助计算偏微分
lenear 无需考虑此问题
Overfitting:A more complex model does not always lead to better performance on testing data.
hidden factor->
back to step1:redesign the model
back to step2:regularizaton
越平滑，对noise越不敏感。training data误差越小，但testing data误差不一定
