NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/comp/ai/neuraln/2933

Internet Message Format | 1992-07-25 | 3.8 KB

Path: sparky!uunet!dtix!darwin.sura.net!mojo.eng.umd.edu!disney.src.umd.edu!tedwards
From: [e-mail] (Thomas Grant Edwards)
Newsgroups: comp.ai.neural-nets
Subject: Re: Thanks Neural Nets, and Brains?
Message-ID: <[e-mail]>
Date: 24 Jul 92 19:27:34 GMT
References: <[e-mail]> <[e-mail]> <arms.711907358@spedden>
Sender: [e-mail] (C-News)
Organization: Systems Research Center, Maryversity of Uniland, College Park
Lines: 62

In article <arms.711907358@spedden> [e-mail] (Bill Armstrong) writes:
>First off, isn't it rather strange that the most widespread artificial
>model of neural operation. the multilayer perceptron, uses continuous
>quantities on its connections, while the dendrites and axons of
>neurons use "zero or one" type action potentials?  Until physiological
>psychologists start studying adaptive logic networks, can anyone
>expect much progress on understanding the brain?

I think this is a simplification of what is going on in brain...

o  There are hundreds of different kinds of neurons in the brain.
   Most have a spiking nature, but I think to consider the spiking
   nature an ALN is not proper, in most situations.  Time-domain
   characteristics of neural firing, pulse width and pulse frequency
   modulation, play parts which we are only beginning to properly
   understand.

o  But the electronic nature of neurons is only one facet of neural
   behaviour.  There is a great dependence on chemical messengers in
   the brain, some of which we understand, most of which we do not.
   Some of these pass from neuron to neuron in complicated ways, others
   just diffuse through the tissue.  There is even talk of microtubules
   providing a significant path for chemical messengers.

o  In many neural systems, we find gap-junctions, which operate like
   analog resistive networks

Anyway, to conclude, the brain is complex, real complex.  We don't understand
it's spatial organization, we don't understand the neuron fully,
we don't understand the supporting cells either.  The major path of
getting a understanding of the brain is to approach it from the senses.
For example, we are just about getting an idea about how the retina works
(which is, more or less, an extension of the brain.  Alot of complex
spatio-temporal processing goes on there).  And we are even becomming able
to trace operation back through the LGN, into the parietal area.  But
that's about as far as our understanding takes us.  Similarly, we are also
approaching brain from the auditory system, trying to grasp the
processing and lateral inhibition across frequency.

We are at the point where we can start trying to study some small
isolated neural circuitry, assuming that neural modulating messengers
are controlled or not playing a large part.  But clearly, there is alot
of biochemical and physiological research left to be done.

What about neural nets?  Well, they are clearly not simulations of
real neural activity.  The study of neural networks, if not taken too
literally, I believe can give us the basis of understanding the massive
parallel computation going on in brain.

For example, I think the extensive work with 3-layer perceptrons has shown
that homogenous assemblies of computational entities probably is not the
best model for explaining the incredible parallel computation in brain.
We can look at Cascade-Correlation as an example of a self-adaptive
architecture that has differentiated modules, yet still maintains a high
degree of parallelism, and it gets alot more jobs done than simple MLP's.
I wouldn't nominate CC as a real brain model, but as a theoretical tool
for understanding how parts of brain might be organized.

I also believe work on chaos in man-made neural nets may also be a
useful tool for understanding chaos in real neural systems.

-Thomas Edwards