New Interpretations of Normalization Methods in Deep Learning