Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!wuarchive!udel!haven!purdue!ccncsu!longs.LANCE.ColoState.EDU!ld231782
From: ld231782@longs.LANCE.ColoState.EDU (Lawrence Detweiler)
Newsgroups: comp.ai.neural-nets
Subject: Re: Postulates on the number of neurons
Message-ID: <11678@ccncsu.ColoState.EDU>
Date: 12 Dec 90 06:53:12 GMT
References: <1990Dec11.174921.26838@cs.columbia.edu> <1990Dec11.040646.20760@noose.ecn.purdue.edu>
Sender: news@ccncsu.ColoState.EDU
Reply-To: ld231782@longs.LANCE.ColoState.EDU (Lawrence Detweiler)
Organization: Engineering College, Colorado State University
Lines: 142


muttiah@welsh.ecn.purdue.edu (Ranjan S Muttiah) in
<1990Dec11.040646.20760@noose.ecn.purdue.edu :

>What do researches think about why there are so many _individual_ neurons
>in the human brain ... ?  Is there some critical number
>needed for "intelligent behaviour" ?

esrmm@warwick.ac.uk (Denis Anthony) replies in
<1990Dec11.152350.16930@warwick.ac.uk>:

>The speed of processing of the nervous system (milliseconds) and the amount
>of computation required for (e.g.) vision in real time, and the number of
>simultanious computation systems (vision, sound, proprioception, cognition
>etc.) require a large number of parallel processes. (all pretty obvious).

Nearly any facet of human perception can be used as a plausible example for
the necessity of the massive, unrivalled parallelism in the human brain,
particularly vision.  Clearly intelligence requires extraordinary
computational power, but parallelism is only one way to satisfy this
prerequisite (although the approach seems more and more a prerequisite
itself!)  Biology, in all its grandeur, is rarely a brilliant technological
innovator in the eyes of one of its offspring, those sole thinking critics
(after all, it hasn't even invented and applied the *wheel* in any of its
abundant creations and variations.)  Parallelism is an unmistakably intrinsic
element in biology, however.  It is the secret of life's success.  Even though
single-celled organisms (bacteria etc.) probably outnumber the multi-celled
ones in our realm, I would wager the vast majority of cells are contained in
multicellular organisms.  So biology has had some bias for parallelism for
some time, and we have something to learn from our patient and pragmatic
engineer.

I'm sure there are many opinions on what comprises the "critical mass" in
human brain tissue that separates them so distinctly from the rest of the
animal kingdom.  One possible candidate is the huge amount of "interneurons"
in the human brain.  Neuroscience found a distinction of motor- vs. sensory-
vs. inter- neuron useful in categorizing the nervous systems of invertebrates.
Motor neurons connect to muscle fibers, or nearly so, and sensory neurons
connect to receptors.  Unfortunately, as we climb the vertebrate ladder, so to
speak, this distinction becomes increasingly meaningless, whereupon in homo
sapiens we reach a climax where only one in many thousands of neurons can be
labelled a "sensory" or "motor" neuron, and the vast remaining majority take
on that nebulous and mysterous term "interneurons".  Therein lies thought,
somewhere.  Some might balk at the idea of some "critical mass" of neuron
number whereupon we reach an explosion of intelligence.  However, the idea is
not so absurd to dismiss without thought...our brains have changed remarkably
little in terms of organization in comparison with other invertebrates.  The
lobes and bulges tend to do the same thing in the same places across species
(the hippocampus and motor cortex come to mind).   We humans just have this
incredibly large bulge in one of the usual bulges, namely the neocortex, that
houses this vast majority of interneurons, and there is a lot of
circumstantial evidence that this is precisely our differentiating advantage.

The "phenomenon" of sign language in apes may be a good example of the role
that the interneuron plays.  It is very intuitive to think that the
interneuron is what facilitates abstraction and association, arguably the
foremost underlying qualities of speech.  Here we may have a very vivid
demonstration of interneuronal power--apes, which have a fraction of the
interneurons that humans do, but still more than other animals, can
communication in few-word sentences.  Perhaps if we grafted additional
associational cortex in their brains the complexity and abstraction of their
"speech" would be enhanced!  (When people criticize studies of sign language
in apes as mimicry, I tend to consider this a manifestation of a superiority
complex, or more precisely an "anthropomorphic" complex.  Why discriminate
against them just because we can do it better?!  We don't scoff at our babies
for their verbal mimicry!  In terms of evolution, apes, just as babies, are
the immature reflections of ourselves.)

What about this "critical mass" of neurons?

>Clearly there is a minimum number, [the order] I do not know. There is some
>redundancy in that fairly large portions of the cerebrum may be lost before
>gross brain damage is manifest (though other parts of the brain are
less robust).

There is a long history of misunderstanding the inherent quality of
fault-tolerance in a massively parallel system.  Reports of its nature have
been greatly misinterpreted.  Because we remove nine-tenths of a rat's brain
and it still navigates a learned maze with minimal impairment, does that imply
that nine-tenths of our mental capacity is dormant?  We are simply applying a
concept where it is not appropriate.  In truth, this concept of "graceful
degradation" is a somewhat novel concept to human engineers.  There aren't too
many engines that gradually slow down as we remove pistons, or buildings that
gradually lean while we remove columns!  Just how crucial is one neuron to a
thousand?  One one-thousandth!

minsky@media-lab.MEDIA.MIT.EDU (Marvin Minsky) in
<4311@media-lab.MEDIA.MIT.EDU>

>It would seem likely, at least to me, that most single cells don't do
>useful operations alone, so that, for example, it might need a whole
>cluster or column of cells to do some little useful job.  So for many
>functions, you'd be left with "only" a few hundred million" "units".
>And it would seem that an ordinary "commonsense database" semantic
>network might well require a few dozen million nodes and connections.
>No one knows that yet, of course.

Every neuron is doing a useful job, its just that we have a difficult time
measuring/quantifying/characterizing its diffuse and miniscule contribution. 
Perhaps we can duplicate some aspect of the effects of some clusters of
neurons with some other device, but we will probably make compromises in the
versatility and plasticity of the system.  Perhaps the ultimate animal, "more
than human", would be the synthesis of evolutionary biology and human
technology.  But there may be some features of the biological approach that
are so ideal (for two, its versatility and plasticity) that they would be
essential to the feasibility of the system.  However, the idea of alternate
forms of intelligence, perhaps to the point of improvement, is a very enticing
idea!  It is the motivation behind the field of AI!

>At the other end of the opinion spectrum are theorists who suggest
>that each single cell might store hundreds, or even millions of bits
>inside the cell.  No one has made a good case for this, or for how
>they might be read-in or -out, but it certainly isn't inconceivable.

At another end of the opinion spectrum are theorists who suggest that concepts
such as  "storing bits" in neurons is meaningless and embarassingly
technocentric.

fahn@cs.columbia.edu (Paul N. Fahn) in
<1990Dec11.174921.26838@cs.columbia.edu>

>It seems to me that (computer-based) neural net researchers too often
>presume that biological organization is somehow optimal. It may be that
>the brain represents an enormous waste of resources, and that 
>"intelligent behavior" can be achieved with far fewer neurons, if only
>they were organized more "intelligently". (or, equivalently, that higher
>intelligence could be achieved with the same number of neurons.)
>
>We should definitely look at biological nets for ideas, but in a very
>critical way, and be prepared to reject anything found in the biological
>realm.

That's an interesting question that takes one step farther Marvin Minsky's
favorite idea that neural networks can be condensed (ideally
computationally)--not only can units be amalgamated, but some can be
eliminated!  Is there a neurological equivalent of an appendix or tonsils? 
Someday we may be surprised and aghast at the depth and vastness of the
evolutionary baggage we carry...  But only when we understand the design we
will have the opportunity to transcend it.


ld231782@longs.LANCE.ColoState.EDU