Normalization before a new prediction?

Course Q&A Machine Learning Specialization Supervised ML: Regression and Classification
1

Hi,
I am currently working on a multivariate linear regression.
I’m a bit confused about the normalization (mean=0 and standard deviation=1). If we do normalization on our training set, shouldn’t we also normalize our cross-validation set as well as all the new values of X that the model does not know?

I have the same question for image classification with a CNN. If we do pre-processing on our training images, shouldn’t we do the same pre-processing on a new image that the model does not know?

Regards,

Pierre

2

Hi!

Yes, every transformation of the data that you do in training should be done on the new data as well.

3

Hi, @Pierre_BEJIAN !

Just one more thing to add to the post. You should normalize your test data with your training data mean and std dev. In a real world scenario you don’t know how your input data is distributed, so you cannot rescale it with any other values than the ones from your training.

4

You are totally right! I had guessed it but it’s good to have confirmation.
Thanks a lot