Midterm Notes for ML2¶

If we pick squared cost we use mean If we use classifiers its best to use mode

c = squared error r = l1 c = cross entropy r = l1 c = squared error + l2 r = l1 (since l2 is smooth we can add it to the smooth part c)

x2 may go zero if we use l1

lower bias because higher k gets closer to "true error" of our given dataset?

larger k increases the covariance between each \(f_{k}\) since they're trained on the same data

prototype is a generalization of rbf kernel