Understanding And Improving Efficiency In Training Of Deep Neural Networks