Path: utzoo!utgpu!utstat!jarvis.csri.toronto.edu!mailrus!cornell!uw-beaver!fluke!ssc-vax!bcsaic!ray
From: ray@bcsaic.UUCP (Ray Allis)
Newsgroups: comp.ai
Subject: Re: Chinese Room Argument
Keywords: understand mind symbol
Message-ID: <10704@bcsaic.UUCP>
Date: 12 Mar 89 09:11:34 GMT
Organization: Boeing Computer Services ATC, Seattle
Lines: 142


 Roel Wieringa points out,

   "It is part of the definition of a Turing machine (TM) that it manipulates
   symbols on the basis of their form only.  The denotation of the symbols,
   and of expressions built from these symbols, is not relevant for the
   outcome of the TM computation."

I, too, have pointed out several times that symbols are deliberately stripped
of their connotations and denotations before they are submitted to "symbol
processing".  Symbol processors *do not* handle meanings, by design.  It is
*specified* that meanings have no place in symbol processing; they get in the
way.  

*After* the processing (predicate calculus, rule-based expert systems,
whatever) *we humans* may attach meaning to whatever symbols emerge.  (I
encounter heavy resistance about here, usually).  There is no meaning carried
through the "symbol grinder"; there is nothing present in the symbol which
determines its meaning.  The meaning of a symbol exists in a mind, in the
association of the symbol with the mind's personal experience.

Drew McDermott says:

   "I hope it's obvious that the rules I and the others are envisioning do a
   good deal more than the Schankian script appliers Searle was describing. 
   I will grant that those rules wouldn't have had experiences."  

It doesn't MATTER how complex the rules are in the Chinese Room, if the room
only receives symbols as input.  It makes no difference how sophisticated the
processing of the symbols inside the Chinese Room, the meanings were all left
outside the door, along with any associations between meaning and symbol. 
None of that enters the process.  When symbols are returned from inside the
Room, some mind associates or assigns meaning to them.  And, fuel for further
discussion, there is *no guarantee* that those associations are "the same" as
any which might have existed before.  

This came from USENET rec.games.chess:

Remarks by M. Valvo re: the US Amateur Team Championship East 2/18-20 1989

    "I was the operator for Deep Thought and I was amazed at its poor play in
	the openings.  At first I thought it was because of its black repertoire,
	but, while that is also true, I realized its understanding of how to play
	openings on its own was nearly non existant."

[I would remove the word "nearly".] 

    "They play perfect tactical chess for anything within their range.  If
    something exists outside their sight, they are helpless."

Regarding openings: "The Alekhine, which Deep Thought seems to like, is
	not a good choice.  It requires concepts.  If you play over the Alekhines
	played by Deep Thought, it is clear that once it is out of the book, it
	flounders around"

After adjustment for the anthropomorphisms, it is apparent that even if chess
programs are winning games at master level, something is missing.  Chess is a
logical system, a classic symbol system, all syntax and no semantics.  The
system is self-contained: there are markers or tokens, and a complete set of
rules for their manipulation.  There is no "denotation" or "connotation"
(required) from outside the system.  It would be susceptible to play by
strict deductive logic if the necessary calculations could only be performed
in less than the remaining lifetime of the Universe.  

We're quite sure that human chess players do not (can not) calculate game
play to anything like the extent computer programs do.  And yet the human
players are still quite competitive; the programs have not yet overwhelmed
humans.  No, humans use metaphors like "get into fire fights whenever
possible".  Humans "understand" the analogy between the volatile battle
called a "fire fight" and a particular approach to move choice in chess. 
Humans use many other analogies and metaphors to play chess.  One side of
each of these comes from human experience.  The meanings of phrases like
"fire fight" are not part of the definition of chess, but the denotations and
connotations of that phrase and hundreds of others are indispensible in the
guidance of human play.  But denotations and connotations are not allowed
into the symbol-processing computer.  

Considering why chess programs and machine translators and "expert systems"
are not very satisfying led me to see that symbol processing is not enough,
by itself, to produce human thinking.  We think with the denotations and
connotations, which is to say our experience.  Symbol manipulation (i.e. 
logic) is a tool which helps us think; it doesn't think by itself.

I think Gary Schiltz's story of passing a calculus course without
"understanding" calculus and the Feynman anecdote are excellent evidence for
the existence of systems which simulate understanding without possessing it. 
Of course those deceptions were not sustainable.  How about

   "The past decade has seen the creation of a substantial number of AI
   programs that are capable of making discoveries at a non-trivial
   (professional) level.  Such programs include Meta-Dendral, AM and EURISKO,
   BACON and its associates (DALTON, GLAUBER, STAHL), and KEDADA."
   
Does any reasonable person believe these programs "understand" anything at
all?

The present argument is whether any device (in this case the "Chinese Room")
using only symbols, can "understand".  Even with a very relaxed definition of
"understand" the answer is (to me) clearly "No".  This conclusion is in
direct opposition to the PSSH.  It says that, if the goal was really human-
like intelligence, decades of work on symbol-only systems has been futile. 
You just can't get there that way.  You can't walk to the moon.

What is at issue here is explanation of the fact that decades of work based
on the Physical Symbol System Hypothesis (PSSH), various Predicate Calculi
and the Clockwork Hypothesis have failed to produce intelligent machines. 
Even if you continue to resist defining terms of discussion such as
"intelligent", "understand" or "common sense", it's hard to deny that fact.

Stevan Harnad talks about symbol grounding, I talk about non-symbolic
analogs, Roel Wieringa points out that symbolic systems are not affected by
denotations or connotations of the symbols.  I can't speak for them, but I'm
questioning some of the fundamental assumptions of AI here, both research and
application.

Now, if you've read this far, I have a question.  But first, some heresi..
er, um, hypotheses:

    Human "natural" languages are not symbol systems; nothing useful can be
	done with only the symbols.  It's the meanings that are important. 
	Directly translating from the symbols of one language (e.g. English) to
	the symbols of another (e.g. Chinese) without recourse to denotations and
	connotations is nonsensical.  (This really isn't arguable, is it?)

    Thinking and understanding have to do with (non-symbolic) physical and
	chemical events in our central nervous system (brain).  Neural nets
	(biologically inspired) are interesting precisely for these non-symbolic
	attributes.  Rules are an after-the-fact description of those events. 
	Rules cause nothing. 

I have tried to present these ideas in other places than comp.ai, and in
general, they are even less accepted than here.  To all confirmed, hard-over
PSSH believers, is there anything that would convince you that there's more
to humans than symbols?  (Something short of an actual, working
implementation please, as that will take some time.)  Is there some evidence
that would cause you to re-inspect your conviction that the "Systems Reply"
is sufficient?   What part of the process by which I came to see the
inadequacy of the symbol processing approach can I explain more clearly?

ray@atc.boeing.com - 

Disclaimer: redundant; my employer has already disclaimed me.