I am facing quite a strange behaviour. As far as I have understood, especially when dealing with CNN, feature normalisation/ standardisation, should help the model at converging faster. Now, I am working with ECG signals, which are characterized by high variability intra and inter subjects. I am dealing with them using CNNs. My model performs quite well when I do not normalise the data. Otherwise, when I do apply either standardisation or normalisation, performance downgrade (higher loss, and 10% less accuracy and F1). This is true if I apply the preprocessing globally as well as if normalize/standardise each signal per se.

Any clue about what could cause this behaviour? My guesses are:

  • presence of outliers that causes too large "compression" of the data with low values
  • as before but with numerical problems caused by the usage of float32, that could cause (?) potential loss of data and thus of information.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.