Matching amplitude to frequency makes it easier for a NN to learn.

AVG loss has been computed as the average of 4 best of 4 training runs, on each target function.