I am confused. Dose get_updates apply the gradient to the trainable variables directly? Or just return the gradient?