Prove that the update step in the large$-$margin kernel perceptron, the inner part of the for$-$loop in the algorithm above, is equivalent to a steepest descent mini$-$ mization of the hinge loss on a single example with respect to a single parameter of the model.