Re: >H soft incremental uploads

Eugene Leitl (Eugene.Leitl@lrz.uni-muenchen.de)
Wed, 14 Aug 1996 16:33:37 +0200 (MET DST)


On Tue, 13 Aug 1996, Anders Sandberg wrote:

> On Mon, 12 Aug 1996, Eric Watt Forste wrote:
>
> [ gobble ]
> > Techniques for manufacturing new designs of computing hardware that can
> > instantiate neural and fuzzy computational processes are developing *much*
> > more rapidly than our understanding of the brain.
>
> Let's see now: Moore's Law: processor power doubles every 18 months.
> Medical knowledge doubles every 5 years. Yes, you seem to be right.
> (although knowledge about the brain grows faster than most areas, thanks
> to recent insights).

Don't fall for one of the oldest IT myths: what is processor power? Is it
the clock or the producer MIPS rating? Hardly so. Is it what a test suite
measures? Better, but then there are lots of ratings, depending on what
is being measured. If you measure local Flops the result will be
drastically different from measuring local or nonlocal memory bandwidth.

Memory bandwith latency progress is asymptotic. Even worse, the price
fluctuates wildly due to market politics. Even the physical clocks will
run into a saturation zone very soon. Afaik Alpha will soon run at 0.5 GHz.
End of line for current Si/Ge semiconductors lies about 2 GHz. And this
is local clock, friends and neighbours. Off-chip accesses are waaaay off.
Trying to run a macroscopic bus at >100 MHz is asking for trouble.

About the only thing which goes up exponential is integration density.
But what shall we do if it runs well into nm domain? Wires don't conduct
at quantum scale too well.

Sorry, system power has rencently started to show signs of saturation.
We'll need to go maspar, soon. The advent of SMP (which is an extremely
poor parallelism) in high-end PCs is a portent for the future.

> > [ yummy ]
>
> If we can create some way to interface at all, we can start hook in
> functionality. The interface will just be very messy and non-intuitive,

Yeah. This is merely going slightly beyond Biomuse.

> unless we can hook up signals to the normal I/O systems (sight, hearing,
> touch) to run our brittle packages (I *want* matlab in my head - I'm

I think it will need heavy training, preferably in a very young age to
make sense of all the extra data. Though to some it may appear horrible,
I think babies should be equipped with implants soon after birth. Only
they can become the true power users.

> falling in love with this old, crusty and powerful package right now).

:))

> > [ attractors frolicking ]
>
> There are some speculation that the attractors replicate as a part of
> mental functioning. In this case adding more "cortex" would give us a
> larger mental space. Let's hope its true! :-)

Yes, but I wish this has happend at the genome level. (Toying with the
size of the cortex is easy: the genome is quite modular). On the other
hand, I remember reading that brain metabolism is even now a bottleneck
at some (early) developmental stage...

> > [ gobble ]
>
> I think most of the brain is rather static in circuitry, but its meaning
> can be very quickly redefined (remember the experiments with glasses
> turning everything upside down - after a while vision worked as normal).
> My guess is that parts of the cortex are undefined from the beginning,
> and later used in various systems that emerge.

Yes, but even then their malleability is limited. Typically, the
topographically nearer the region, the more likely is it to take over the
function of the affected area.

'gene