On the neural approach for FET small-signal modelling up to 50 GHz