X-Message-Number: 9272
From: Thomas Donaldson <>
Subject: Re: CryoNet #9262 - #9267
Date: Fri, 13 Mar 1998 00:48:05 -0800 (PST)

Hi everyone!

As Art Quaife pointed out some time ago, we will never work with a true 
Turing machine, but only finite versions of one. And already it's true that
a Turing machine, if set to computing quite a number of results (such as
the weather, or even the configuration of a complex molecule) would take 
far longer than we would be willing to wait. The point of the counterexample
is not that this is a BETTER machine, it is merely not a Turing machine. Mr
Freeman was good to bring up the actual reference, but he should not so 
easily denigrate it on the grounds that the computer devised would be
too slow. Ditto for Turing machines. So there.

Even before Siegelmann came up with that example I have been pointing out,
right here on cryonet, that parallel computers raise some fundamental issues.
Sure, you can imitate them with Turing machines, but you would need very great
patience to do so. Not only that, but they differ in one quite essential
way (anyone who has tried parallel programming will know this like they 
know the back of their hand, but for others it may come as new): you do NOT
want recursive computations. The problem with recursive computations is
that they are necessarily sequential; if you want to attain the same aim,
you work out a nonrecursive algorithm that gets you to the same place.
This tells me that recursive algorithms and computers, even theoretical
ones like the Turing machine, miss out on something essential about 
the world. After all, we live in a world in which lots of things are
going on simultaneously, and not just independently, but they are responding
to one another at a high rate. And someone decides to present a model of a
computer for which working out such processes becomes very long and 
difficult, and takes a long time ... too long. Something is very wrong
here. (Yes, I know who those someones were, but that does not move me to
emend what I've said).

Whatever our brains do, they too operate in parallel, essentially. Every
single neuron is a quite complex little machine, responding to many different
chemicals and producing an equally wide number of outputs. For some special
cases that evolution has produced in us, we are very fast, not because
our chemistry is fast but precisely because we are parallel, not sequential.
For that matter, our neurons are ANALOG computers (if that notion applies
at all). It's that feature which made me wonder whether or not we might 
do the same things that Siegelmann's counterexample does. Remember that
Siegelmann did not present a LARGE neural net, just a small one, and 
parallelism can produce results very fast if you have lots of processors
connected together in the right way. This was more a question than an
assertion when I suggested it, and I still think it remains a question, 
most especially because we have no reason to believe that our axons and
dendrites are multiples of some fixed small length.

For Mike I will have this to say: Mike claims that if we could store 
ourselves in a computer then it would follow that we could make that
stored person run in a computer. Here is the hiatus in his reasoning: we
do not just run autonomously, EVER. We are always responding to something
in the world, and incommensurability is essential in the world. In the
first place, it would take far more processing power to run a human being
in a computer than simply to have the human being there in the world, 
responding and acting. Secondly, whatever equations describe our behavior,
they are highly likely to NOT converge to some single solution, while
the computer, because it is digital (I believe we still need a good 
definition of that, Mike) can only produce some values of the person,
not all those a real person would produce if that person were in the
world. Running someone in a computer would ultimately fall into the
same resolution problems that computer graphics does: sure, you can get
finer and finer resolution, but since the world is not made up of small
distinct locations with given colors etc. your picture will always have
limits below which it becomes quite false. Why is that important? Because
even small differences will increase with time until they become quite
large --- that is why it is so hard to predict the weather, even with
the best supercomputers. It is even the same with detailed predictions of 
the motion of planets: the farther back you go, the finer become the 
computations needed. Doing the same with a human being, I believe,  
will be many orders of magnitude harder... so much so that doing such
a computation would require far too big a computer to be practical, even
in the far future.

Any theory of computation based on sequential computing, discrete sets of 
events, and recursive algorithms will ultimately fail to accurately deal
with the world. And why try to do that in the first place?  

And for those who may decide that I'm making arguments against computers
or whatever, I will add that these ideas in no way prevent us from 
constructing intelligent beings/machines/devices/people. What they are
saying is that the wrong way to do so is to try to do it with digital
computers; start making your devices up out of analog computers, to 
start, and not just one but lots of them, even if you're only trying
to build a creature with the intelligence of a paramecium. The brain
(and you'll need more than just a brain!) you produce will consist of 
several very large neural nets, possibly of designs not yet devised, 
with millions of nodes. It will have some special circuits that produce
a single awareness, usually occurring after most of the processing has
completed, and of course apparently sequential to the brain and to 
others (all the massive parallelism remains unconscious by the construction
of this brain). And so you can make a device very like a human being. And
the fact that all the nodes in this highly parallel machine are analog
rather than digital is essential to its design.

Of course I'm glad we have computers, and would even say that constructing
an entire working person (by other than the means now used!) might not
be a good idea ... though making parts and using ideas from an understanding
of how we work will probably be very useful. Neural nets and fuzzy computing 
have started to become very useful, just as examples. Computers can do some 
things for us much better than we do them ourselves; most simply, do 
calculations. Evolution has not adapted us to do rapid calculatons. And so
we augment ourselves with tools, just as we always have. But these tools are so
useful precisely because they CAN'T make a human brain. If they could, we'd
find them quite useless for our calculations, terrible as database engines,
and disobedient as robots.

So that's what I have to say to my various readers. And thanks for bringing
up the reference. I really would like to take up this issue again when I
am in Australia, as I said before.

			Best wishes, and long long life,

				Thomas Donaldson

Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=9272