Neural Network border condition or extrapolation data

Question

I was reading the user guide of NN for Matlab and I found this quote about extrapolation data:

It is important that the data cover the range of inputs for which the network will be used. Multilayer networks can be trained to generalize well within the range of inputs for which they have been trained. However, they do not have the ability to accurately extrapolate beyond this range, so it is important that the training data span the full range of the input space.

My concern is about testing (or validating) some observation located near the outside boundary that probably increase the error. So in that cases I was thinking don't evaluate this observations. By the way, in my problem I have few data so the model it's more sensitive to outliers.

Someone know an article or some explanation about this?

Answer 1

My concern is about testing (or validating) some observation located near the outside boundary that probably increase the error. So in that cases I was thinking don't evaluate this observations.

This is a recipe to give your estimates of error an optimistic bias. Systematically eliminating these data from your test sets means you're only measuring the performance of the model on the data that's most similar to the training data. The purpose of the train/test split or cross-validation is to examine how well the model does against unseen data.