Path: utzoo!utgpu!water!watmath!clyde!att!osu-cis!tut.cis.ohio-state.edu!mailrus!ncar!boulder!sunybcs!bingvaxu!leah!gbn474
From: gbn474@leah.Albany.Edu (Gregory Newby)
Newsgroups: comp.ai
Subject: Re: Intelligence / Consciousness Test for Machines (Neural-Nets)???
Summary: my $0.02 (what is this intelligence thing, anyway?)
	 (a self concept based approach to intelligence)
Message-ID: <1120@leah.Albany.Edu>
Date: 18 Oct 88 06:38:37 GMT
References: <1141@usfvax2.EDU> <3430002@hpindda.HP.COM>
Organization: The University at Albany, Computer Services Center
Lines: 79


(* sorry about typos/unreadability:  my /terminfo/regent100 file
 is rapidly approaching maximum entropy)
In article <3430002@hpindda.HP.COM>, kmont@hpindda.HP.COM (Kevin Montgomery)
writes:
> In article <1141@usfvax2.EDU>, mician@usfvax2.BITNET writes:
> > When can a machine be considered a conscious entity?
> 
> May I suggest that this discussion be moved to talk.philosophy?
> 
> While it has many implications to AI (as do most of the more
> philosophical arguments which take place in comp.ai), it has a
> broader scope and should have a broader reader base.  

I would like to see this discussion carried through on comp.ai.
It seems to me that these issues are often not considered by scientists 
working in ai, but should be.  And, it may be more useful to
take an "operational" approach in comp.ai, rathar than a philosophical
or metaphysical approach in talk.philosophy.

This topic has centered about the definition of consciousness, or
the testing of consciousness.

Turing (_Mind_, 1950) said:  "The only to know that a machine is
thinking is to be that machine and feel oneself thinking. " (paraphrase)

A better way of thiking about consciousness may be to consider
_self_ consciousness.  That is, is the entity in question capable
of considering its own self ?  

Traditional approaches to defining "intelligent behaviour" are
PERFORMANCE based.  The Turing test asks a machine to *simulate* a human.
  (as an aside:  how could a machine, which has none of the experience
  of a human, be expected to act as one.  Unless someone were to 
  somehow 'hard-code' all of a human's experience in some computer system, 
  but who would call that intelligence?)
Hofstadter (_Goedel, Escher, Bach_, p24) gives a list of functions as
criteria for intelligent behaviour which many of today's smart expert
systems can perform, but they certainly aren't intelligent!

If a machine is to be considered as "intelligent," or "conscious,"
no test will suffice.  It will be forced to make an argument on its
own behalf.  

This argument must begin, "I am intelligent"

  (or, "I am conscious" --means the same thing, here)

The self concept has not, to my knowledge, been treated in the AI
literature.  (My thesis, "A self-concept based approach to artificial
intelligence, with a case study of the Galileo(tm) computer system,"
SUNY Albany, dealt with it, but I'm a social scientist.)

As Mead (see, for instance, _Social Psychology_) suggests, the
difference between lower animals and man is twofold:

1)  the self concept:  man may consider the self as an object, separate
from other objects and in relation to the environment.

2)  the generalized other:  man is able to consider the self as
seen by other selves.

The first one's relatively easy.  The second must be learned through
social interaction.


So, (if anyone's still reading)
What kind of definition of intelligence are we talking about here?
I would bet that for any performance criteria you can give me, if I 
gave you a machine that could do it, the machine would not be considered
intelligent without also exhibiting a self-concept.

'Nuff said.

--newbs
  (
   gbnewby@rodan.acs.syr.edu
   gbn474@leah.albany.edu
  )