X-Message-Number: 8595
From: Thomas Donaldson <>
Subject: Re: CryoNet #8552 - #8560
Date: Sun, 14 Sep 1997 16:58:04 -0700 (PDT)

Hi guys, once again!

One point which really stands out: at no time have I criticised Alan Turing
himself, or his ideas. And yes, he knew some math. I was criticising some
later interpretations of him and his work by others.

When Mr. Merel says that a Turing machine should be able to imitate a 
neural net, I will even agree with him (modulo the question I raised in 
my last posting). However I see no way to make it imitate a neural net
while taking the same TIME TO DO SO as the neural net. A different notion
of equivalence occurs here.

As for Mr. Minsky, I would certainly agree that he is an intelligent man
and his book contained intelligent ideas. HOWEVER my response to it comes
from reading the book, which begins with the claim that it will tell how
our brains work and ends with the claim that just maybe they work as
described. In between, we basically see the ideas of someone familiar
with computers but not with brains. Some of those ideas are good ones,
some look weak after 20 more years of work. But my editorial was called
"Ptolemaic Science": and even Ptolemy had a good deal of useful stuff
to say about astronomy. We use his ideas still for rough calculations 
about where to look for something. That hardly means that he got the 
Solar System or the Universe anywhere near right.

Further, a careful study of brains does NOT mean that you look only at
the neurons. You get to look at how they are connected, and use PET
scans and other ways to see which ones are active, and so on. In animals
(yes, we assume that brains of birds and mammals work in extremely
similar ways to our own brain) you can do even more. My statement was
NOT a statement about levels of investigation but about the hardware
itself --- and if you don't like to think of brains as hardware, then
OK, you can call it squishware without changing what I'm saying.

Finally, about future devices: no, I was not trying to be difficult
or unfair. I really don't claim to know just what kind of devices 
we will use in 2400 to do some of the things we do now with computers.
Perhaps we'll have working quantum computers (incidentally I was NOT
suggesting that our brains differ from present computers because of
some issues of uncertainty). Perhaps we'll use very long molecules for
data storage (that's almost been tried, in fact, and may become useful
for special cases quite soon). If someone wants to be uploaded into 
a computer, not for storage but to actually remain there as a working
being, then it is clearly important for them to think out just what
they mean by a computer, uploading, and get a better idea of just
how such a system might work.

There are other issues raised by Mr. Clark. The very first point I will
make is that a computing device using floating point numbers rather
than digital numbers, and organized a bit like a Turing machine, is 
NOT my original idea. I got it from reading the computer science 
literature, and thought it interesting, too. (PERIASTRON didn't get
that report, mainly because it was pretty far from the subject and aim
of my newsletter --- I don't print everything. I will hunt for the 
originators and where they published once I get several other things
off my plate. I believe it was in one of the 1997 issues of either
SCIENCE or NATURE. Its inventors, as you might guess, actually
proved that some devices of this kind could not be imitated by Turing
machines --- and not at all for practical reasons, but for the same
kind of reasons Turing invented his conceptual machines). 

Again, Mr. Clark somehow believes that I said the difference between
neural nets and computers was qualitative. Perhaps I said so, I don't
remember: but qualitative and quantitative shade off into one another.
Neural nets are highly parallel, and even a mouse brain has more neurons
by orders of magnitude than any parallel machine yet built for computing.
If we use a standard for equivalence which ignores issues of the length
of time required for calculations, then yes, neural nets are equivalent.
If we require that calculations be done on more or less the same time
scale, then they are not at all equivalent. If that is a qualitative
difference, then we have one.

			Best and long long life,

				Thomas Donaldson

Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=8595