Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.milw.wisc.edu!lll-winken!uunet!portal!cup.portal.com!dan-hankins
From: dan-hankins@cup.portal.com (Daniel B Hankins)
Newsgroups: comp.ai
Subject: Re: Question on Chinese Room Argument
Message-ID: <16872@cup.portal.com>
Date: 8 Apr 89 20:04:50 GMT
References: <10992@bcsaic.UUCP>
Organization: The Portal System (TM)
Lines: 104

In article <2691@crete.cs.glasgow.ac.uk> gilbert@cs.glasgow.ac.uk (Gilbert
Cockton) writes:

>For one, human minds are not artefacts,

     Webster's Dictionary defines an artifact as an "object made by man." 
I will grant you this assertion for the nonce, merely noting that this is
in fact subject to considerable debate.


>whereas computer programs always will be [artifacts].  This alone will
>ALWAYS result in performance differences.

     This I can dispute.  Many programs are written not by humans, but by
other programs.  Some are even written by the environment.

     Take the example of a neural network implemented in hardware.  This is
definitely a computer, of the massively parallel variety.  The program here
is the combination of the set of activation levels of the neurons and the
set of neurochemical levels of the synapses, as well as the interconnection
topology of the neurons and synapses.
     This computer's program is _not_ made by man.  It is made by the
environment.  This environment may include the influences of humans, but
nevertheless the program is not man-made, any more than a human mind is.

     Now, consider a neural network program running on a computer with an
unspecified number of processors and duplicating the behavior and internal
structure of the hardware version.
     The neural network program is made by man, certainly.  It is an
artifact.  But there is another level of program here; the same one
mentioned previously.  The activation levels of the neurons and so on form
_another_ program, one that is again made by the environment rather than by
man.

     Genetic algorithms are more examples.  The GA itself is man-made, but
the genotypes it generates are not.  Those genotypes form a program for the
behavior of the phenotype, whether explicitly in the form of machine
instructions, or implicitly in the form of parameters.

     As a matter of fact, any program that is written by another program or
by a combination of another program and the environment cannot be
considered to be man-made, and therefore cannot be considered an artifact.

     Arguments that the ultimate cause of the end-program is human through
the mechanism of the neural network hardware or software or GA apply
equally to human minds; human minds have as their ultimate causes the
mechanisms of conception, gestation, and education.


>Given a well-understood task, computer programs will out-perform humans. 
>Given a poorly understood task, they will look almost as silly as the
>author of the abortive program.

     You must be thinking of expert systems, inference engines, and other
rule-based systems that do not model themselves and do not modify their
rules based on experience.

     Control of chaotic systems such as pipeline flow control are poorly
understood tasks.  So are many problems in visual recognition.  Yet ANNs
and GAs are quite adept at solving these problems - sometimes much better
than their authors.

     I know nothing about how pipeline flow control works.  But, given a
particular setup, I could write a GA to do it very well in a few days.


>The issue as ever is what we do and do not understand about the human
>mind, the epistemelogical constraints on this knowledge, and the
>ability of AI research as it is practised to add anything at all to
>this knowledge.

     Perhaps that's what you consider the important issue.

     I wasn't aware that the only aim of AI research was to expand our
knowledge of the human mind.  I thought that there were some other goals,
such as producing useful software, playing God (making beings in our own
image), "because it's there", and so on.
     One does not always build a fire in order to learn more about the
chemistry of combustion; sometimes it's important just to stay warm. 
Conversely, knowledge of the chemistry of combustion is not needed in order
to start a fire.


>Come on then boys and girls in AI, lets hear it on "suitable" :-)

     You're on.  I'd say that a suitable program would be one that is
self-modeling, capable of both generalization (ANN-like) and deduction
(inference-engine-like), pattern recognition, pattern association, and many
others I can't think of at the moment.

     We're a long way from a computer program and sensor/effector hardware
that can achieve sentience - but we have some good ideas of what directions
to go in.  The connectionist direction (if the hardware were there to
support it on anything like the scale in the human brain) seems to be one
of the most promising to me.


Dan Hankins

     A group of men in the Garden of Gethsemane were engaged in an odd
     activity.  One spun a wheel and called out a number.  The others
     studied parchments they held, and one cried, "Bingo!"  The
     wheel-spinner smiled.  "Don't write this down, John," Jesus said. 
     "This is part of the _secret_ teachings."