From: Ben Goertzel (ben@goertzel.org)
Date: Tue Sep 16 2003 - 22:08:36 MDT
> >If you ever feel interested in doing some collaborative thinking
> on how to
> >structure your mega neural network in a useful way, I'm sure
> that both Peter
> >Voss and I would be very willing to toss around some ideas together with
> >your team. While we don't have hardware infrastructures on the scale of
> >yours, we've been thinking about how to structure quasi-neural computing
> >systems for quite some time (although my own AI design Novamente, isn't
> >really all that quasi-neural anymore, it did start out that way).
>
> We do welcome the collaboration with other researches. In fact, that was
> partially the porpuse of our presentation: To reach groups like yours :-)
>
> Best Regards,
> Marcos Guillen
Although my Novamente system is NOT neural net based, I've thought a lot
about how to make a neural net based AI system.
The reason I didn't choose the NN path is that it seemed to me it would be
*much* more consumptive of computing resources than the Novamente approach.
I.e., it seemed to me that the NN approach relied on a lot of redundancy.
But on the scale you're working, you can afford a fair bit of redundancy, so
you can probably do some interesting things.
Specifically, I've figured out some modifications to the standard hebbian
learning (long-term potentiation) rules, which seem to lead to neural nets
whose behavior approximates probabilistic logic on an "emergent level" (i.e.
on the level of bundles of connections between neural clusters). The
dynamics of the average weights of the connection-bundles over time will
often approximate probabilistic logic in a certain sense. So it's a cell
assembly model that seems to give rise to symbolic logic in an emergent
sense. At least, to certain parts of logic -- I haven't yet pursued that
train of thought to its conclusion to see how far it can be pushed.
So, when you get to the point that you've got your big-ass neural net
running and your problem is making it learn, it would be interesting to talk
to you about these "hebbian logic" learning rules and their possible
applicability in your architecture.
This is not something I'm currently actively working on; and nor have I
published it yet, more due to lack of time to write it up carefully than due
to any desire for secrecy...
-- Ben G
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:42 MDT