Is normalization necessary for input and output data in an Artificial Neural Network (ANN)
No, it is not necessary.
Does it impact the network's performance?
Yes, it does affect network's performance.
When features are to different scales, then methods such as gradient descent tend to favour the features unequally. This results in effectively weighting some features differently than others. If you are considering each feature as having value relative only to itself, then you need to normalize in order to be fair. If, though, the values have meaning between features, then you do not always want to rescale.
