Xref: utzoo comp.ai:5480 sci.philosophy.tech:1890
Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!umich!samsung!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!sdsu!ucsd!sdcsvax!mandrill!bloch
From: bloch@mandrill.ucsd.edu (Steve Bloch)
Newsgroups: comp.ai,sci.philosophy.tech
Subject: Re: more Chinese Room
Message-ID: <7693@sdcsvax.UCSD.Edu>
Date: 11 Jan 90 08:41:27 GMT
References: <2602@cunixc.cc.columbia.edu> <1527@skye.ed.ac.uk>
Sender: nobody@sdcsvax.UCSD.Edu
Reply-To: bloch@thor.UUCP (Steve Bloch)
Organization: University of California, San Diego
Lines: 116

pnf@cunixa.cc.columbia.edu (Paul N Fahn) writes:
>                                       He is basically asking people to
>try to identify with a computer cpu and then conclude non-understanding.

Let's give Searle this, for the moment: the CPU does not understand,
by analogy with the Anglophone in the Chinese room.  It's still the
wrong question, since Searle said he would disprove that "computers
running programs cannot understand", a completely different
proposition.  Nobody's ever seriously claimed that the processor
itself somehow magically becomes intelligent when a certain program
is stuck into its memory; the claim (if we are to trust Searle's
statement of it) was that if a suitable computer, running a suitable
program, passes a Turing test, then it actually is intelligent.
The analogy of the Chinese room is invalid unless it includes the
whole system: Dr. Searle in his room, the book of rules (or just "the
rules", if Dr. Searle somehow manages to memorize enough formal rules
to completely specify the verbal behavior of a human being, which I
suspect exceeds the theoretical information content of a human brain),
and the window through which Chinese characters are passed.  And the
question of whether this system "understands" Chinese is no easier
than whether an AI program "understands" fairy tales.

jeff@aiai.UUCP (Jeff Dalton) writes:
>The point of the Turing Test is not to answer whether a computer does
>or does not understand but rather to substitute a different question
>which we may find good enough.
I'm sure Searle, and for that matter many AI researchers, would agree
with you: the Turing Test isn't sufficient to prove that a suitable
computer running a suitable program understands.
But he goes farther: his objective is to prove that a computer 
running a program CANNOT understand, although some other unspecified
kind of machine doing something else unspecified might.  His certainty
on this issue seems to me quite unwarranted, and sometimes downright
offensive.

Paul goes on:
>Let us look at Searle's recent twist to the problem: the man memorizes
>the rules and answers Chinese questions in public. We the experimenters
>watch him do this for two years and then must decide whether he 
>understands Chinese. A lot of people would conclude that he does indeed
>understand Chinese, even if they "knew" that he was following rules.

To which Jeff replies:
>No they wouldn't, because they'd sooner or later ask him to restate
>some Chinese in another language.  Besides, you're assuming that a
>lot of people would accept a Turing Test as adequate.

Don't we?  I'm using the Turing Test to conclude that both Paul and
Jeff are people, and I have no qualms about it.  Indeed, if Dr. Searle
were on this newsgroup I suspect he would do the same.

As for "they'd ask him to restate some Chinese in another language",
if they asked in Chinese, he would presumably answer in Chinese that
"I don't speak English."  If they asked in English (equivalent to
hitting BREAK on your terminal, falling into a ROM monitor, and
talking to the processor in machine language), this would require
throwing out part of the system that was passing the test, so the fact
that the new system fails indicates nothing about the old system.

********************************************************************

Another objection to Searle's article.  Searle makes much of the
distinction between a model and an actual object:

"a person does not get wet swimming in a pool full of ping-pong-ball
models of water molecules,"
"Simulation is not duplication."
"you could not run your car by doing a computer simulation of the
oxidation of gasoline, and you could not digest pizza by running the
program that simulates such digestion.  It seems obvious that a
simulation of cognition will similarly not reproduce the effects of
the neurobiology of cognition."

When you talk to someone on the phone, you are not actually hearing
the person's voice, but rather a simulation of it, carried out by
formal, mechanical processes, with a fair amount of digital signal
processing and no small amount of random noise introduced to boot.
Yet for practical purposes we treat this simulation as the reality,
because what we're interested in is communication of ideas, not the
physical presence of a person within earshot.  (In some cases there
isn't even a person at the other end, just a machine simulating a
person's voice straight into the wire.)  In other words, the
simulation preserves the part of reality we're interested in.
If I have an idea, it remains essentially the same idea whether I
speak it aloud, type it on a keyboard to store it in an ASCII file on
a magnetic disk, or write it in Chinese with a horsehair brush,
inkstone and inkblock on rice paper.  These representations are all
simulations of one another which preserve the essential part of the
idea, the real information.
By contrast, when you ask whether I've gotten wet, you're asking about
interactions between my skin and water molecules; the ping-pong-ball
model does not preserve interactions with real skin, unless it too is
modelled on the same scale.  When you ask whether a car has moved,
or whether I've digested a pizza, you're asking about the release of
energy (among other things), and a computer simulation of oxidizing
hydrocarbons (or carbohydrates) doesn't preserve the release of real
energy.

"Simulation is not duplication" is certainly true unless you have a
COMPLETE simulation of EVERY ASPECT of the simulated system.  But an
incomplete simulation CAN duplicate the aspects of the system you're
interested in, and if what we're interested in in cognition is
information flow, there's no reason to believe a computer program 
can't simulate that (it is their specialty, after all).
And while it may "seem obvious that a simulation of cognition will
similarly not reproduce the effects of the neurobiology of cognition,"
I don't care because neurobiology isn't what I'm interested in.
Searle, on the other hand, apparently takes it as axiomatic that
cognition cannot occur without certain biochemical reactions.
Everything depends on what aspects of cognition you care about:
the Turing test is perhaps a little overly behavioristic, but I
think Searle's demand is at least as far in the other direction.

"The above opinions are my own.  But that's just my opinion."
Stephen Bloch
bloch%cs@ucsd.edu