Xref: utzoo comp.ai:8229 sci.bio:4173 sci.psychology:3884 alt.cyberpunk:5354
Path: utzoo!attcan!utgpu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!usc!apple!portal!cup.portal.com!mmm
From: mmm@cup.portal.com (Mark Robert Thorson)
Newsgroups: comp.ai,sci.bio,sci.psychology,alt.cyberpunk
Subject: The Bandwidth of the Brain
Message-ID: <37034@cup.portal.com>
Date: 18 Dec 90 07:02:23 GMT
Organization: The Portal System (TM)
Lines: 93

There is a common myth that the brain is capable of enormous computational
bandwidth -- for example that the retina sends gigabauds worth of data to
the brain.  I believe the computational bandwidth of the brain is quite low,
low enough that we could simulate a brain on today's computers if only we knew
how to do it.

As the first piece of evidence, consider the results of applying information
theory to animal behavior.  It may come as a surprise to hear that zoologists
have developed statistical techniques for measuring the bandwidth of
communication between animals.  This is possible because animals (especially the
lower species, such as fish and insects) have very stereotyped behavior
patterns.  In experiments I performed using crickets, these behaviors consisted
of waving the antennae, fluttering the wings, chirpings, and a few other things.
I would put a cricket in a box with a little matchbox house.  Crickets like
little enclosed spaces, so the cricket would go into the house and make it its
home.  After some time, I would introduce a second cricket and record the
interaction as the first cricket defends its nest against the interloper.

The interactions were very consistent.  First there would be an antenna wave
followed by an antenna wave from the intruder, then there would be a wing
flutter or a chirp, etc., finally resulting in the defender chasing off the
intruder.

I would compile data for dozens of such interactions.  To convert this data
to an estimate of the bits exchanged between the animals, I organized the
data in a matrix, with stimulus events in the columns and the response
events in the rows.  Then there was some statistical technique I've forgotten
for boiling down the matrix to a single number representing the number of
bits exchanged.

What was surprising was just how few bits are exchanged when animals interact.
In my experiments, only about 2 or 3 bits were being transmitted per interaction.
The professor of the course had a table summarizing many experiments with
other species, showing a rise in information transfer as you go up the scale
to humans, who (by this measure) can assimilate hundreds of bits per second.
This seems to jibe with reading speed -- I can almost read text blasted at me
at 1200 baud, which seems about the highest-bandwidth input that I have.

Another piece of evidence comes from reaction time tests.  These are performed
using an instrument called a tachistoscope, which is a rear-projection screen
upon which images can be flashed.  By simply asking you to respond when you
see a number flashed on the screen, we can get a figure for the speed of the
path from the eyes through the brain to the muscles.  Then, by changing the
experimental paradigm -- for example, by flashing simple math problems -- we
get a longer reaction time.  The difference between the two times is the
amount of time the brain needs to do the additional work.  By dividing this
number by the speed of nerve cells, it's possible to make an estimate for
the number of stages of nerve cells which were involved in performing the
task.  (See _Human_Learning_and_Memory_ by Roberta Klatzky for a good
introduction to the topic.)

What was surprising was how few layers are involved.  Even fairly complex
math or word-association tests seemed to correspond to 10 layers or less.

So it seems like the whole brain, engaged in a task which undeniably involves
thinking, might be modeled as a pipeline of 10 stages with no more than
1200 baud bandwidth each -- an astoundingly low amount of computational bandwidth.
Of course, this doesn't mean the same 10 stages are used for every problem,
merely that most sorts of thinking don't involve many layers of cells or
much bandwidth.

I think the reason people believe the brain has enormous computational bandwidth
is that people see bundles of nerve fibers, and assume they are like wires
in a computer or a communications network.  They falsely assume that each fiber
is an independent channel, and that the total channel capacity is the product
of multiplying the capacity of an individual fiber by the number of fibers.
This is clearly not true -- you can't have all the fibers in your spinal cord
jumping simultaneously.  Likewise, when I view a bit-mapped graphics display
with my retinas, I cannot simultaneously perceive all the dots on the screen
and I certainly can't remember or interpret them if they are changing 15 times
a second.  Presented with a single frame of random dots, I might be able to
memorize some small 10 x 10 grid subset of the image if given enough time
to memorize them (like an hour).

I think it is obvious that the brain consists of many agencies which are
"on call", but very few agencies which are simultaneously active.  Our 
remarkable ability to remember millions of minor facts, or recall some
minor event which occurred many years ago (and which one hasn't thought
about for many years) is no more remarkable than the ability of a phone book
to record the names and addresses of millions of people or the ability of
the disk drives at TRW to store millions of credit histories.

The evidence suggests that the parts of the brain that are active
during any short span of time provide very low computational bandwidth;  their
power comes from the fact that many such parts are available to be used.
I don't use my math parts while making dinner, I don't use my cooking parts
while writing this Usenet posting.  And I haven't used my Morse code parts or
my German-language parts much at all in the last 20 years.

Existing computers have far more computational bandwidth than is needed to
simulate a human consciousness.  What is needed is a model which allows the
parts that are "on call" to reside on disk and the parts which are active
to be accessible in semiconductor memory.