Is possible to normalize just the training data instead of the whole data using Neural Network toolbox?

Using the Matlab functions, it is possible to normalize the input and output data that will feed the Neural Network. However, is possible to just normalize the training data instead of the whole data set? Is this a better approach than normalizing the whole data?

Answer 1

[Greg Heath]

The normalization parameters ARE obtained from the training subset; then applied to all of the data.

So, you have no problem!

Hope this helps.

Greg

[Anderson]

Hi Greg,

I have seen that using the same data and random division of the training, validation and testing sets I always get the same parameters for the [-1,1] mapping (mapminmax) in the generated neural network file. It sounds that the mapping mapminmax is doing it for the whole data instead of just for the training set. Is there a way to fix this?

Answer 2

[Greg Heath]

No.

Why would you want to do it?

It doesn't make any sense to me.

Greg

[Anderson]

Hi Greg,

I have read that, in order to have a general neural network, the normalization parameters should come from the training set instead of the whole data available. Is it correct? Or is the normalization of the whole data better? Why?

Thank you!

[Greg Heath]

Reread my answer.