[From Bill Powers (970203.0400 MST)]
On Saturday I saw a show, "Discovering Women," on PBS. It featured the
career (so far) of Mischa Mahowald. While it was partly a study of how a
woman gets along in a career where men predominate, it was also about a
group of researchers at Cal Tech (and Oxford) who are modeling the brain as
an analog computer, on VLSI (very large-scale integrated circuits --
millions of transistors), silicon chips which they design and have built.
They call what they do "neuromorphic engineering." Misha has participated in
the design of a "silicon retina" and has made important -- dazzling --
contributions to building an optic nerve to link retinas to central
computing machinery, and has also made circuits that can use information
from dual retinas to generate analog depth information about objects in the
visual field. Her PhD thesis was defended on May 12, 1992: it can be
downloaded from
www.pcmp.caltech.edu/anaprose/mishathesis.pdf
Gsview (which links to and uses Ghostscript) can read pdf files, or you can
get Adobe Acrobat free. Warning -- the file is a megabyte, and the thesis is
248 pages long. There's also a postscript file, but it's twice as long and
Netscape warns that it can't understand the coding ("x-gzip"). I read the
whole damned thing last night, until my eyeballs were spinning. Didn't
follow it in detail, of course, but if I had to I COULD! What a sensation.
I've been trying to locate Misha Mahowald on the internet; so far she seems
to be a "former member" of the neuromorphic engineering group, and while
there is a Web address given for her at Oxford (zen.pharm.ox.ac.uk), it
wasn't answering the phone on Sunday. There's still a mailbox for her at
Caltech, but I don't know if she gets messages there.
The head of the lab at CalTech is Carver Mead. The home page for his group
is www.pcmp.caltech.edu. It's worth exploring some of the links, including
reports on an annual meeting in 1996 at Telluride, CO., with, presumably, a
1997 meeting coming up. In the report, it's evident that some real progress
is being made in modeling the circuitry of the brain, and that it's just
what is needed in PCT. One of the subgroups is concerned with sensory-motor
behavior, and they're on the wrong track, so maybe this group needs PCT, too.
Rick, have you heard of this group? It's under the "Computation and Neural
Systems Program" at Caltech in Pasadena. Right under your nose. Robert
Kosara, you mentioned a method of neural modeling in which only meaningful
events were transmitted -- is this Mahowald's optic nerve method? I sounds a
lot like it. This is a digital device, but Mahowald remarked on the TV
program that she did it that way because it's hard to model the growth of
axons by an analog method -- a digital method is the easiest way to set up
connections, I guess. But all her work (with Mead) on the retina and her
method of depth perception uses strictly analog methods. Mahowald also
mentions in her thesis that the analog methods are the only way to go to get
the necessary computing speed: digital computers (even supercomputers) are
way too slow. She points out in her thesis that using transistors (and
neurons) just to represent on and off conditions is a horrible waste of
their computing capacity.
By the way, Robert, the chapter on "Premises" in B:CP is one that my editor
at Aldine really screwed up. Originally, it just went right into the subject
of neurons as analog computers, but he thought the readers would drop like
flies that that point unless I put in a long teaser about what would be next
in the book after all the boring stuff. So that's why there is all the
premature stuff about levels at the beginning of that chapter. I didn't want
it there, but he insisted. I'm sure you will be able to see the exact spot
where the digression starts.
I'm going to call Carver Mead today, to see if there's any glimmer of
interest there, and also to find out where Misha is now.
There really is a world, apparently, where people actually still understand
analog circuitry and analog computing. The digital approach has so
overwhelmed the modeling community that I'm really amazed, not to mention
excited. No more "states!" No more "iterations!" The analog circuits work
exactly as they should, everything happening at once, in parallel,
continuously. Wow. Robert, are you part of this world?
Oh, I wish I were young again (but not the same as when I was young, of course).
Best to all,
Bill P.
I found two more URLs: