Block-Normalized Gradient Method: An Empirical Study for Training Deep Neural Network