Blackman Window

Hello,

neural networks have IMO a tendency to lead to "alchemy". Nothing wrong
with NNs as such, but they usually don't offer any insight to the problem.
And some folks (like me) have become quite sceptical over all the hype.

You could repeat the sine wave experiment at a much lower and higher
frequency (let's say 50 Hz and 2000 Hz). If the results don't match,
you're dealing with memory effects, meaning the model requires knowledge
of the past signal.

(Almost) foolproof techniques to deal with memory effects -and- weak
nonlinearity at the same time have been known for many years. Search for
"Hammerstein model", "Wiener model" or "Volterra series" in general.
"Strong" nonlinearity is a bit harder to deal with, but if it doesn't work
for the former, it won't work with the latter either.

Hint: for a thesis it may be a good idea to compare NN results with what a
run-of-the-mill polynomial etc model delivers.
A result that NNs aren't any better can be a very respectable result AS
LONG as you've understood both :) 

-mn

 

Sorry, but, correct me if i'm wrong... you mean, in order to check if
i'm dealing with memory effects also, i should run the network (which
has been trained with a sin at 400 HZ), with a new sinus signal
(50-2000 Hz) ?.

It's a great idea.... thanks for the tip!!!.

I'll also look up for Hammerstein, Winer and Volverra. Thank you!!!!!!

Regards

It's been something I wanted to try for quite a while:

http://www.dsprelated.com/showarticle/19.php

-mn