Normalization source for neural network during application (not training)

I've built and trained a neural net using z-score normalization of the input parameters. When I now go to apply that network I again apply z-score normalization, but i'm unsure what source to normalize with....the mean and standard deviation of the training data, or the mean and standard deviation of the new data? It seems like a simple question, but I couldn't find a clear answer.

Answer 1

[Greg Heath]

A basic assumption is that the design (training and validation) data adequately characterize the summary stationary statistics (e.g., means, variances, maxes, mins and correlations) of all nontraining data (validation, testing and unseen).

The defaults of random datadivision and validation stopping are included to help prevent poor performance on nontraining data.

Now, if new data is encountered whos summary stats vary significantly from the design data, then

a. It could be included in a new design

b. It could be used to obtain an alternate design.

Hope this helps.

Thankyou for formally accepting my answer

Greg

P.S. Don' forget that plotting is a very effective way to recognize nonstationary and incompatible statistics.

[Shannon]

Thanks Greg, if I understand you correctly you're saying that as long as the summary stats of the new 'applied' data are comparable to those used during training and validation then I should use the summary stats of the training data in doing z-score normalisation of my applied data.

Regarding checking that, at current I'm performing PCA and reviewing that the new applied data is at least within the limits of the PCs representing 99% of the initial data variability. Does that sound reasonable to you?