Path: utzoo!attcan!uunet!lll-winken!ames!mailrus!uflorida!haven!aplcen!jhunix!ins_atge
From: ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards)
Newsgroups: comp.ai
Subject: Re: Question on Chinese Room Argument
Summary: internal vs external sensory
Keywords: Water, Sensory-Motor I/O, internal reps
Message-ID: <917@jhunix.HCF.JHU.EDU>
Date: 1 Mar 89 16:44:00 GMT
References: <45126@linus.UUCP> <5662@homxc.UUCP> <45199@linus.UUCP>
Reply-To: ins_atge@jhunix.UUCP (Thomas G Edwards)
Organization: The Johns Hopkins University - HCF
Lines: 42

In article <45199@linus.UUCP> bwk@mbunix.mitre.org (Barry Kort) writes:
>Recall the breakthrough scene in the Helen Keller Story.
  ...
>The Chinese Room is like Helen before her moment of epiphany.
>There is little point in manipulating symbols mechanistically
>unless one can map the symbols to non-symbolic sensory
>information from the external world.

     It is certainly true to assume that sensory perceptions from
"the enviroment" (outside the cognitive device) are neccesary for
real-world reasoning.  (Galileo pointed that out in _Dialogue
Concerning the Two Chief World Systems_, although he was only talking
about human brains).  Internal rules are not enough.

     However, in the Chinese Room experiment, we are assuming that
the rule operator has indeed been endowed with rules to operate
on, and as such these rules are defacto sensory input from the
outside.  
     Moreover, the incomming Chinese is also sensory input.
Rules may exist which change due to incomming Chinese.

Furthermore, what are these rules?  Do these rules include sensory
information (i.e. is there a rule which deals with what-is-trees
which includes a picture of a tree)???

One more angle on this entire situation is that neural-networks can
often be described by symbollic rules.  Often discovering these
rules from learned weights can be difficult, but there have been
some breakthroughs (i.e. shading --> concave or convex object?
Sejnowski has taught a NN to do the shading to concave or
convex and determined symbolic rules).
   I feel though that NN's give symbollic rules a richer
"spectrum", and that it's much easier to induce new nerual
weights than to induce new symbollic rules.
   Even if you are still not quite convinced that NN's can be
reprsented by symbollic rules, take every neuron to be a rule
which takes the weighted sum of activations of rules connected
to it and perform a function on the rule activation, and propogate
that along directed edges to other rules...

-Thomas Edwards
ins_atge@jhuvms bitnet