NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/comp/ai/neuraln/2919

Text File | 1992-07-24 | 3.7 KB | 78 lines

Newsgroups: comp.ai.neural-nets
Path: sparky!uunet!gatech!destroyer!ubc-cs!unixg.ubc.ca!kakwa.ucs.ualberta.ca!alberta!arms
From: [e-mail] (Bill Armstrong)
Subject: Re: Neural Nets and Brains
Message-ID: <arms.711986585@spedden>
Sender: [e-mail] (News Administrator)
Nntp-Posting-Host: spedden.cs.ualberta.ca
Organization: University of Alberta, Edmonton, Canada
References: <[e-mail]> <[e-mail]>     <arms.711907358@spedden> <[e-mail]>     <arms.711935064@spedden> <[e-mail]>
Date: Fri, 24 Jul 1992 14:03:05 GMT
Lines: 65

[e-mail] (Bill Skaggs) writes:

>You provoked this argument with a claim that ALNs are the only
>reasonable approach to understanding the brain.  I don't dispute that
>ALNs are very powerful and very interesting, but that claim is hubris
>of the worst kind.

I don't fel quite comfortable with the way you phrased that.  My point
was that BP nets use continuous signals and the brain doesn't -- an
obvious very significant difference.  I was asking why people would
expect to understand the brain by studying a system (BP) that is
*different* at the most basic level of signalling.

Don't you agree that if the brain works on 0-1 signals, then to study
the brain one could beneficially look at logical systems? Would you
have accused Galileo of "hubris of the worst kind" to propose using a
telescope to study the planets?  Of course not -- you would try to
understand the merit in the argument for using that kind of instrument.

(If you were saying that I had claimed or implied that the learning
algorithms of ALNs and the brain are in any way related, then I think
your statement about hubris might be quite appropriate, though.)

  The argument from theoretical superiority is
>simply irrelevant.  Mother Nature is an engineer, not a theoretician,
>and the brain is a collection of kludges -- some of them more elegant
>than we have any right to expect -- put together by trial and error.

Sounds reasonable to me.

>I'm not a big fan of backprop, but it does have the advantage of
>simplicity to counterbalance the disadvantage of inefficiency, and
>there is no way of knowing *a priori* which of these Mother Nature has
>considered more important.

Someone was claiming logic networks can't be trained like BP networks
can.  I pointed out that ALNs *are* trainable, by a process that
is very simple.  If it were the case that BP nets can be trained and
logic networks can't, then BP nets would be the obvious thing to study.
But that's not the case.

Along the lines of your argument: BP is gradient descent which is an
obvious thing to try.  ALNs are also "kludges" that go beyond gradient
descent, based on years of experiments.  They still contain a component
of gradient descent though.

As for efficiency: inefficiency may not penalize a brain like it would
penalize a computer -- neurons are cheap and redundant computation is
potentially useful in case of damage.  ALNs are quite different in this
regard -- they are designed for efficiency and avoiding any reundant
computation.

In sum: to study the brain, in my opinion, based on ignorance of how
the brain works, is that one should study a system that works on
logical signals.  Although ALNs might help in this, they are not
designed as a brain model at all, and different learning algorithms
would have to be developed to capture the learning properties of
neurons.

Are we getting closer?
--
***************************************************
Prof. William W. Armstrong, Computing Science Dept.
University of Alberta; Edmonton, Alberta, Canada T6G 2H1
[e-mail] Tel[phone] FAX [phone]