• Choose which features to include
  • Choose the right value for lambda
  • These are what discussed in "mode selection"

  • Overfitting really not a good example for predicting the future examples

  • Suppose we choosing which parameters and which degree of polynomials
  • Here's what it works
  • it could determine how well and generalized the value of d
  • test each d by using jtest(theta) , inject the parameters into the Jtest
  • choosing whichever has the lowest error value
  • Still not a good point for generaliz e the parameters
  • Perhaps is the best, but not always the generalize. May tends to be overfitting.
  • To address the problem, instead of splitting into two training and test, split into three
  • the final is in right corner set is Mtest



  • Use CV to set the model of the test
  • These (d = 1) is the linear model
  • First, extract the parameters, and test that into cross validation (cv) set
  • The lowest cross validation error (this case, d= 4) then use this degree model to the test set
  • Then use the result of error test to the generalization model that we selected
  • Orr, use the model picked to the test set, and estimate generalization error.



  • Jcv will generally be slower than Jtest, because we get parameters from training set, and select the model of parameters from cross-validation.
  • We choose the lowest error of Jcv, from  Xcv & Ycv, so that's why it fit cross-validation, more than test set, which in turn lower cost than test set.
  • Most machine learning practioners use just training set and test set. Its better to have 3 separate set, that is training set, cross validation set, and test set