NetNews Usenet Archive 1993 #3/NN_1993_3.iso/spool/talk/abortion/57409

Text File | 1993-01-23 | 6.1 KB | 123 lines

Newsgroups: talk.abortion
Path: sparky!uunet!news.encore.com!jbates
From: [e-mail] (John W. Bates)
Subject: Re: Myelin (Was Re: Spoken Like a True ProLifer)
Organization: Encore Computer Corporation
Date: Fri, 22 Jan 1993 19:05:58 GMT
Message-ID: <[e-mail]>
In-Reply-To: [e-mail]'s message of Fri, 22 Jan 93 03:58:02 GMT
References: <[e-mail]>
    <[e-mail]>
    <[e-mail]>
    <[e-mail]>
Sender: [e-mail] (Usenet readnews user id)
Nntp-Posting-Host: pinocchio.encore.com
Lines: 106

In article <[e-mail]> [e-mail] (Mark A. Cochran) writes:

>>Sorry, Mark, but you've got things a little mixed up. In networking
>>lingo, hypercubes are a network designed for supercomputing, with
>>a node at each vertex connecting to each neighboring vertex. It's not
>>related to neural networks at all. The closest neural network design
>>I can think of is James Anderson's "brain state in a box", in which
>>each output pattern is a vertex of an n-dimensional box. Nice model
>>for associative memory, but n tends to have to be very large (in
>>computational terms) for it to be useful.
>>
> It's a good thing I made sure to deny any real knowledge of
> hyper-cubes then, isn't it? :)
> I still bet you can't build one out of C=64's though.... :)

Oh, you could build one, all right. It might even be more useful than
a single C-64. That's not saying much. (Actually, on a side note, my
father still uses the C-64 he bought in 1982. He really likes it, but
soon might be upgrading to a 286. Whoooeee! (And besides, look at what
Chaney accomplishes with his C-64...))

>>I've been leary of bringing models into this discussion, since it is
>>often hard to relate models of neural networks to actual neural
>>networks. But now, let me refer to a model by Stanislas Dehaene and
>>Jean-Pierre Changeux, which simulated the performance of human
>>infants in Piaget's A not B task. Their results approximated the
>>performance of human infants.
>>
>>The interesting part of the experiment, though, was that they varied
>>the amount of "noise" that the network received. At high levels of
>>noise, the network performed at the level of a 7-month old infant,
>>but at low levels, it performed at the level of a 12-month old
>>infant.  Noise levels seemed to correspond to the development of
>>myelin in the frontal lobes.  (from the _Journal of Cognitive
>>Neuroscience_ 1:3, S. Dehaene and J.P.  Changeux, A simple model of
>>prefontal cortex function: delayed response tasks)
>>
> Interesting. If they were able to jump the noise level around to
> approximate various developmental stages, I wonder if they could/did
> jump it up to a level that would approximate that development of, say,
> a 22 week fetus?
> Be interesting to see the results if they did. It could shed some
> light on this subject, at least.

Well, the problem with models is that it is entirely possible to read
too much into them. The strongest conclusion you can make from the data
gathered is that the performance of the network matches the performance
of a child at a specific age. Since the mechanism by which the network
befores does not map one-to-one with the mechanism by which the child
performs, it's still only an approximation. Extrapolating the performance
of the network beyond the data we have from actual experimentation is
really only guesswork.

Of course, we can determine the amount of noise in an unmyelinated vs.
a myelinated axon, by using the cable equations. We know that the
performance of a network will degrade rapidly when the noise increases
and/or when the size decreases. But, without experimental data, we can't
be sure that our simulated network matches the actual performance of the
human mind.

> [Crabs and Squid deleted, since I've already had lunch]

>>>>>> The resonable work in question, though, is thought. Just as you
>>>>>> can't use a 4 bit 16K RAM computer as an effective file server, I
>>>>>> don't see how you cna use the similarly limited abilities ofthe
>>>>>> pre-myelinated neural system as a 'thought server'.
>>
>>>>>I'm reserving my judgment on the matter for a time when we know
>>>>>more about all the issues involved.  The hypomyelinated mice
>>>>>discussed later in the post may be our best window into this issue.
>>>>>In the meantime, it must be obvious to everyone reading this thread
>>>>>(all 3 of us :-) that neither of us has any clue about whether
>>>>>myelin is necessary or not.  I think that there is at least a fair
>>>>>amount of information on the biological side that suggests that
>>>>>myelin is not as central as some claim.  On the other hand, your
>>>>>arguments about the presumed complexity of the network are
>>>>>certainly thought-provoking (there's that smell again...myelin
>>>>>burning or something).
>>
>>Yes. The major problem that we have in modelling brain processes is
>>the complexity of the whole thing. I mean, our supercomputers have
>>problems with 2-3000 neuron models. Massively parallel systems
>>reach the 16-32000 neuron level. How much of the brain is actually
>>dedicated to thought? Maybe what, 10^10 neurons?
>>
> I recall reading we'd need a computer the size of (something like)
> Manhatten to approximate the brain. Does that sound like a resonable
> size?

Well, I just did some quick calculations, and came up with over three
hundred million of the new Thinking Machines CM-5 (I think that's the
model.) Stacked three high, that's a solid block about three miles long
on each side. But then you need power supplies, cooling systems, and lots
of cabling.

I suspect (in fact I'm sure) that this model is a bit overkill. Sounds
impressive, but really overly useful. A whole SPARC processor for each
neuron? Phooey. In practice, I think that Intel(?) has a prototype
neural network chip that simulates about 400 connected neurons in parallel.
Assuming liberal space requirements for each chip, that's only a block
7/10 of a mile on each side, and twenty feet high. Of course, there are
a few minor implementation problems to be worked out...

Gotta run.

John