Understanding Training Dynamics of Deep ReLU Networks