SV: Capacity of the brain

Eugene Leitl eugene.leitl at lrz.uni-muenchen.de
Wed Sep 15 15:16:12 EST 1999


Tim Tyler writes:

 > : that's why any 'normal' nervous system will always 'kick the butt' of any
 > : formerly-'normal' nervous system that has anything shoved in-there.
 > 
 > But this is hardly right.  A coprocessor to perform mathematical
 > operations could massively outperform neurons on the same task.  Basically

It depends on the complexity. As soon as you need a lot of atomic
operations (=require nontrivial complexity) wetware starts to
outperform you because it does things mostly via massively parallel
mappings.

You'd need a room filled with most current DSPs burning MW of power to
achieve what the bee does within a cubic mm on a uW budget.

 > we can design the guts of such a co-processor relatively easily, while we
 > have no idea how to go about building one out of wet neurons in any
 > reasonable timescale.
 
If you don't know how design them, GA-grow them.

 > Your point about signal distances does not apper to be relevant - a
 > maths co-processor could take up very little space.  /Perhaps/ the

Relativistic latency is always relevant at the high end, regardless of 
scale.

 > space would be better occupied with neurons - but no-one knows how to go
 > about building a "neuronal" maths co-processor, while a silicon one can
 > at least be envisaged today.
 
If you don't know how to do it it doesn't mean anyone else doesn't.

 > Also, silicon substrates are inherently better than brain material at
 > performing large numbers of sequential calculations rapidly.  If you have

Connectionist architectures are not particular to any substrate. You
can very well go neuronal with buckytube logic.

 > a problem that *requires* fast, serial computations, a neural net is

Fortunately the number of problems which must be done in above way is
negligeable. Humanity has a strange bias towards sequential stuff,
this trend will be reversed within the next decades.

 > simply not the right sort of architecture to use - but it is practically
 > the /only/ type of circuiry you can build from neurons.
 
Biological neurons != artificial neurons.

 > Kurzweil has machines supassing human brain material somewhere around the
 > middle of the 21st century.  I should think that - according to him - it
 > will simply be more efficient use of the available space to scoop out the
 > bulky, slow, neurons and replace them with modern nanotech crystalline
 > computational hardware ;-)
 
Sound idea, unfortunately pretty tough to do in vivo.

 > : if folks go at it on an evolutionary 'time' scale, will they 'get lucky'?
 > 
 > : nope, 'cause, during all that 'time', 'normal' nervous systems'll 'just' be
 > : getting better... there's a bit of Xeno in-there :-)
 > 
 > This doesn't necessarily follow either.  Normal nervous systems can only
 > evolve at the rate of the species that owns them.  By contrast simulated
 > evolution can go through hundreds of generations in a second on a single
 > computer for some tasks.

I don't think anybody actually proposes to do anything with wetware,
but the time gene plumbing can do that it will be made obsolete by nano.
---



More information about the Neur-sci mailing list