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: <16880@cup.portal.com>
Date: 8 Apr 89 20:18:19 GMT
References: <10992@bcsaic.UUCP>
Organization: The Portal System (TM)
Lines: 91

In article <15122@bellcore.bellcore.com> srh@wind.bellcore.com
(stevan r harnad) writes:

>My position was that subjective meaning rides epiphenomenally on the
>"right stuff," and the right stuff is NOT just internal symbol
>manipulation, as Searle's opponents keep haplessly trying to argue, but
>hybrid nonsymbolic/symbolic processes, including analog representations
>and feature-detectors, with the symbolic representations grounded
>bottom-up in the nonsymbolic representations. One candidate grounding
>proposal of this kind is described in my book.

     Aha!  Light dawns!  All this time we've been talking about different
things when we say 'symbol'.

     When _you_ say symbol, you mean a _linguistic_ symbol - a word.  When
_I_ say symbol, I mean any of the organized patterns that a computer works
with, including the analog ones you describe (floating-point numbers).

     Of course a program that works only with English (or Chinese, or
whatever) words cannot achieve or have sentience; it has no sensory
objects to associate with its linguistic objects.

     However, a program that contains sensory input objects (such as sight,
sound, touch, and so on) should be programmable to associate certain of
those words with certain of the sensory inputs (in a fuzzy, overlapping and
recursive way, of course).  This program should then understand what it is
talking about;  its symbols are grounded in sensory data.

     The confusion and arguments were arising from the fact that many of us
were speaking of symbols as relating to any kind of object, not just
linguistic ones.

     Now, we are left with two questions:  Is any thing that passes even
the most rigorous LTT sentient for all practical purposes, and if a machine
passes the LTT, how could it achieve that state.

     I think that the answer to the first question is definitely yes.  The
LTT is the only real test we have for sentience that is not physically
adjacent.  I apply the LTT every day when I converse over the net.  It
would be absurd to assert that those I correspond with might not be
intelligent simply because I can't open up their heads and see if their
brains are organic; the LTT is clearly sufficient to decide.  Whether
Stevan Harnad is an instance of a program or is human or is a cyborg from
the Lesser Magellanic Cloud, I don't know.  But I think I can conclude
pretty easily, _from discussions on the net_, that Harnad is sentient. 
Biological inspection is not necessary.

     There are two possible answers to the second question.  I shall show
that only one of them is conceivable, which will put me in the same
position as Harnad or nearly so.

     1. A machine could achieve sentience by being directly programmed to
        be so.

        Well, if the only interface we are going to give it is the
        linguistic one, we are going to have to program it with enough
        knowledge to use the linguistic link for future learning - we are
        going to have to teach it language and install a complete set of
        past sensory experiences for it to draw on when deducing and
        inducing from its linguistic input.

        I think it is pretty clear that providing a machine with a full set
        of sensory experience up to the point of language fluency (8 human
        years worth, or so) is an impossible task.  We simply won't ever
        know enough of anyone's experience to provide this.  Not to mention
        the gargantuan task of data entry of all the experience and the
        equally gargantuan task of building all the right associations
        between the words and sensory experience.  It's just not practical.

     2. A machine could achieve sentience by being given the proper
        self-organizing properties (say a large and biologically accurate
        self-configuring neural network), a sufficiently rich set of inputs
        (videocam vision, microphone hearing, pressure-plate touch, and so
        on), and a sufficiently rich set of outputs (robot arms, mobility,
        speech generation and so on)

        Such a machine would only need to be given rudimentary, instinctive
        knowledge, some hardwired behaviors (such as jerking a limb away
        from an excessive heat source), pain, pleasure, and a few other
        capabilities, and it would pick up the rest on its own.  Then, for
        the LTT, one isolates the machine from the judge and allows only
        communication via tty.  This is the option I think might work.

        Of course, producing a 'tabula rasa' sentient being by means of
        physical labor and a 9 month manufacturing process is likely to
        remain more economical for some time to come.


Dan Hankins

"                                                 " -