Statistical Learning Theory of Deep Neural Networks: A Generalization Viewpoint