From: Eugene Leitl [mailto:eugene.leitl@lrz.uni-muenchen.de]
> Ramez Naam writes:
>
> > Yes, if de Garis' neural connections are just random,
> there's no way he's
> > going to achieve human-level intelligence, even with 1000x
> the neurons.
>
> GAs are not random. I do not get his manual block assembly
> part, though.
Having read more of his site (http://www.hip.atr.co.jp/~degaris/) I see now
that he's applying GA to breed neural network modules of 100 neurons. So
clearly this is not totally random. (Though he doesn't state what the
fitness criteria are.)
Unfortunately he provides almost no information on how the modules are
interconnected. According to http://www.cs.usu.edu/~degaris/cam/index.html
"These modules are assembled into humanly defined artificial brain
architectures". Apparently de Garis believes that he and his team can come
up with architectures that produce their desired results, though it's hard
to share his confidence without understanding how they plan to do this.
> > 2) If he can really simulate 10^12 neurons in 2011, then
> that leaves the
> > door open for scanning a human brain and replicating its
> architecture in de
> > Garis' neural simulation. In 2011 presumably the only
> option would be a
> > destructive scan, but even this would be phenomenally
> ahead of schedule.
>
> On what evidence do you base this absurd optimism?
I'm actually rather pessimistic on this front. I doubt that de Garis can
simulate 10^12 neurons of sufficient complexity in real-time by 2011. (In
fact, his whole site leaves out the question of runtime speed. They seem to
view memory size as the fundamental constraint.) But if he could, then it
would seem that he would have provided sufficient computational power to
simulate a human brain at the neural level. Hence, the door would be open
to replicating the structure of a human brain in de Garis' simulation and
observing the results.
This archive was generated by hypermail 2b29 : Thu Jul 27 2000 - 14:02:45 MDT