Hinge-loss without margin: Suppose that we modified the hinge-loss on page 179 by removing the constant value within the maximization function as follows:

J =

n



i=1 max{0, (−yi[W · X T

i ])} +

λ

2

W2 This loss function is referred to as the perceptron criterion. Derive the stochastic gradient descent updates for this loss function.