On Mon, Jul 11, 2005 at 10:31:56AM +1000, Stathis Papaioannou wrote:
> Perhaps, perhaps not. For one thing, in the brain's case we are relying on
> the laws of chemistry and physics, which in the real world are invariable;
> we don't know what would happen if these laws were slightly off in a
A systematic error or noise beyond the homeostatic capability of the
simulation would generate nonsense, of course.
So, stay below the error threshold.
> simulation. For another, we do know that tiny chemical changes, such as a
> few molecules of LSD, can make huge behavioural changes, suggesting that
> the brain is exquisitely sensitive to at least some parameters. It is
So, don't put LSD in the simulated brain. Don't zap the CMOS junction with
electrostatics. Don't put the system nearby a Co-60 source. Do not mutate
bits randomly. Do not change the meaning of a primitive randomly every few
ticks.
If it hurts, don't do it.
> likely that multiple error correction and negative feedback systems are in
> place to ensure that small changes are not chaotically amplified to cause
> gross mental changes after a few seconds, and all these systems would have
> to be simulated as well. The end result may be that none of the cellular
Of course. And your point is?
> machinery can be safely ignored in an emulation, which is very far from
> modelling the brain as a neural net. I may be wrong, and it may be simpler
Strawman, again.
> than I suggest, but as a general rule, if there were a simpler and more
> economical way to do things, evolution would have found it.
Biological tissues are not evolved to e.g. work with EM radio, or electron spin for
information processing, or nuclear fission for power sources, or an enzyme to
deposit diamond. Regardless how many gigayears you spend evolving, this will
never be discovered due to kinetic blocks, fitness crevices, and sterile areas
in fitness space which can't be crossed incrementally. Human design doesn't
have that limitation. We can in principle do whatever evolution can do (by
explicitly invoking the process, in an accelerated model), and more.
The fitness function of discrete information processing in solid state is
entirely different from CNS. Most of what the genome does is not devoted to
neural information processing, and, frankly anisotropically excitable
nonlinear medium is a control paradigm from hell.
There are simpler and more economical ways to do things, and we'll be there
in about 20-30 years. Meanwhile, biology reigns supreme in crunch/Joule,
integration density, error tolerance and a few other things, but we're
gaining on it rapidly.
--
Eugen* Leitl <a href="http://leitl.org">leitl</a>
______________________________________________________________
ICBM: 48.07100, 11.36820 http://www.leitl.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE
Received on Mon Jul 11 2005 - 05:41:31 PDT