Hi,

Pablo wrote:

>man, if that simulator only works with binary neurons than its use is even
>more limited than I thought! the rprop algo's I was talking about works
with
I used the wrong words.  The SNNS is very flexible with a variety of
activation functions available and the option for user defined functions as
well.  I expect it would take me until summer or so to make much progress on
this idea.  If you have code you would like to share that would be great.

My original thoughts were along the lines of Richard Dobson's where the
network would be trained on separate system and then once trained would be
used in Csound as sort of a Plug-In.  As far as training I could perhaps
load up some lightly used workstation and run it over the weekend or just
let it run in the background for a few weeks or however long it takes.

An interesting idea about having the network learn how to create patches to
duplicate a sound.

The SNNS file format is not to great for reading into Csound.  I would
prefer a simpler text format.  In that case a Perl script might be written
to convert from SNNS to Csound NN format.  Other converters could be written
for other NN file formats so people could develop on other systems.

The source code is available for the SNNS however it cannot be modified and
distrubuted.

Regards,
Hans Mikelson

Bye,
Hans Mikelson