Why do we need to do normalization or standardization
Let us explain a word is to the process of solving the loss of the lowest value in more steady and slow, easy convergence.
Detailed explanation can be seen here:
feature project "normalized" What is the role? - Answer Yi Zhen - known almost
https://www.zhihu.com/question/20455227/answer/197897298
With my own words conclusion is: the loss of function when a weight requirements gradient will use the variable before the weight, if the white can be excessive, will lead gradient is too large, that is, the weights convergence process each change will have a significant amount of time, convergence is not easy.
This requires that we use different learning rates in the convergence process, how to use gradient little learning ratio, a large gradient changes direction using larger learning rate, which is too much trouble.