New Patch

[Alp Mestanogullari]

Don't worry too much for the speed for the moment, the code (of the core)
will go through an optimization phase soon, I would like some fusion to
happen there if possible.

I will check out your patch. Thanks Thomas !

On Mon, Mar 8, 2010 at 12:25 PM, Thomas Bereknyei <tomberek at gmail.com>wrote:

> Added a new patch to hnn.hs.
>
> It has two new functions:
>
> step - This advances the example NN a single timestep.  This allows all
> sorts of cycles and anything.  The added complication is that one either has
> to keep track of where the "wave" of signals is, or simple iterate it enough
> times for it to settle.  Bringing us to the next funciton
>
> runXtimes - This simple repeats a endomorphism on NN a certain number of
> times.
>
> *The simplicity of runXtimes is promising, but step involves so much
> unpacking, mapping, and "busy work code" that I think there are issues with
> how NN is constructed*.  Nevertheless, I think I'm on the right track.
>
> To compare this with the previous versions I've put out: It is slower, but
> allows more types of NN's.  This is because each neuron and edge can have
> its own activation and weighting functions, and those lookups are somewhat
> expensive.  The clarity isn't there yet, but I'll keep thinking and trying.
>
> -Tom
>
> _______________________________________________
> Hnn mailing list
> Hnn at projects.haskell.org
> http://projects.haskell.org/cgi-bin/mailman/listinfo/hnn
>
>


-- 
Alp Mestanogullari
http://alpmestan.wordpress.com/
http://alp.developpez.com/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://projects.haskell.org/pipermail/hnn/attachments/20100308/1d3837aa/attachment-0001.htm