How wiggly should the model be?

Filled dots train the model. Hollow dots are new data it has never seen.

3
Training error0
Error on new data0
Just right

Complexity is the degree of a polynomial fitted to 14 noisy points from a smooth curve. Errors are mean squared error.