polynomial regression: regularization vs overfitting
— true
— Ridge
— Lasso
— unregularized
λ
0.01
degree 10 polynomial, 15 points
constraint geometry: why L1 produces sparsity
— L2 circle
— L1 diamond
— loss contours
drag the optimum
constraint size
1.0
coefficient paths as λ increases
Ridge
Lasso
λ
0.00
8 features, synthetic data