Question 10

[SeanV]

Originally Posted by yaser

Correct. The solution was also given in slide 11 of Lecture 12 (regularization).

yes my point was how do you solve this numerically - given that people will already have a good least squares code ( doing SVD on Z to avoid numerical ill conditioning), there is no need to implement (poorly) a new regularised least squares solver

you can just add a few data points at the end of your training data and feed it into your least squares solver. ie
\lambda |w|^2 = \sum_i (y_i-\sqrt(lambda)w_i)^2

ie if w is d dimensional you append to your Z matrix the additional matrix=sqrt(lambda)*eye(d) and append a d vector of zeros to your y

(eye(d) is d by d identity matrix) [ but this is much better explained in the notes i linked to]