🚀 go-pugleaf

It's never clear to me how seriously to take these comp.ai discussions,
but if there really is anyone out there who is interested in a deeper
analysis of the recent Scientific American article by the Churchlands,
here's one. These were comments on an earlier draft, but no changes seem
to have been made, so they apply to the published version too. For the
record, although his Sci. Am. paper was not the most cogent version
of Searle's position, I don't think the Churchland rebuttals work,
so Searle's Argument comes out on top again.
-----
To:    P & P Churchland
From:  Stevan Harnad
         
    THINKING THE UNTHINKABLE, OR, RUN THAT BY ME AGAIN, FASTER

Hi Pattie and Paul:

Thanks for sending me your Scientific American draft. I've seen
Searle's companion draft too. Here are a few comments:

(1) Unfortunately, in suggesting that Searle is the one who is begging
the question or assuming the truth of the hypothesis that syntax alone
can consititute semantics you seem to have the logic reversed: In fact,
Searle's the one who's TESTING that hypothesis and answering that question;
and the Chinese-Room thought-experiment shows that the hypothesis fails
the test and the answer is no! It is the proponents of the "systems
reply" -- which merely amounts to a reiteration of the hypothesis in
the face of Searle's negative evidence -- who are begging the question.

By the way, in endorsing the systems reply in principle, as you do
(apparently only because of its counterintuitiveness, and the fact that
other counterintuitive things have, in the history of science, turned
out to be correct after all), you leave out Searle's very apt RESPONSE
to the counterintuitive idea that the "system" consisting of him plus
the room and its contents might still be understanding even if he
himself is not: He memorizes the rules, and henceforth he IS all there
is to the system, yet still he doesn't understand Chinese. (And I hope
you won't rejoin with the naive hackers' multiple-personality gambit at
this point, which is CLEARLY wanting to save the original hypothesis at
any counterfactual price: There is no reason whatsoever to believe that
simply memorizing a bunch of symbols and symbol manipulation rules and
then executing them is one of the etiologies of multiple personality
disorder!)

As to the speed factor: Yes, that is one last holdout, if it is in
fact true that Searle could never pass the Chinese TT in real time. But
that's at the price of being prepared to believe that the difference
between having and not having a mind is purely a function of speed!
The phenomenon of phase transitions in physics notwithstanding, that
sounds like a fast one, too fast for me to swallow, at any rate.
Besides, once he's memorized the rules (PLEASE don't parallel the speed
argument with a capacity argument too!), it's not clear that Searle
could not manage a good bit of the symbol manipulation in real time
anyway. The whole point of this exercise, after all, is to show that
thinking can't be just symbol manipulation -- at ANY speed.

I don't know about you, but I've never been at all happy with
naive hackers' claims that all there is to mind is the usual stuff, but
(1) faster, (2) bigger, and (3) more "complex." I think the real
punchline's going to turn out to be a good bit more substantive than
this hand-waving about just (a lot) more of the same...

(2) All your examples about the groundlessness of prior skepticism
in the face of physical theories of sound, light and life were (perhaps
unknowingly) parasitic on subjectivity. Only now, in mind-modeling, is
the same old problem finally being confronted on its home turf. But All
prior bets are off, since those were all away-games. The buck, as Tom
Nagel notes, stops with qualia. I'll show this specifically with the
example below.

(3) It is ironic that your example of light = oscillating
electromagnetic radiation should also hinge on speed (frequency). You
say that Searle, in a parallel "simulation," would be waving the magnet
much too slowly, and would then unjustly proclaim "Luminous room, my
foot, Mr. Maxwell. It's pitch black in here!" But here you see how all
these supposedly analogous forms of skepticism are actually parasitic on
subjectivity (with shades of Locke's primary and secondary qualities):
Because of course the only thing missing is the VISIBILITY of light at
the slow frequency. It made perfect sense, and could have been pointed
out all along, that, if fast electromagnetic oscillations really are
light, then it might only be visible to the eye in some of its
frequency ranges, and invisible but detectable by other instruments in
other frequency ranges.

That story is perfectly tenable, and in no way analogous to Searle's
Argument, because it is objective: It's not "what it's like to see
light" (a subjective, "secondary" quality) that the Maxwellian
equation of light with EM radiation is trying to explain, it's the
objective physical property that, among other things, happens to be the
normal cause of the subjective quality of seeing light. The mistake
the sceptics were making is clearly NOT the same as Searle's. They
were denying an objective-to-objective equation: One set of objective
physical properties (among them the power to cause us to see light) was
simply being shown to be the same as another set of objective physical
properties. No one was trying to equate THE SUBJECTIVE QUALITY OF LIGHT
ITSELF with something objective. (Not until lately, that is.)

So, whereas concerns about subjectivity might indeed have been the
source of the earlier scepticism, all that scepticism was simply
misplaced. It was much ado about nothing. Ditto with sound and life:
Subjectivity, though lurking in each case, was really never at issue.
As Nagel puts it, one set of appearances was simply being replaced by
(or eliminated in favor of) another, in the new view, however
surprising the new appearances might have appeared. But no one was
really trying to replace APPEARANCES THEMSELVES by something else, by
the stuff of which all appearances would then allegedly be made: THAT
would have been a harder nut to crack.

But that's the one modern mind-modeling is up against, and Nagel is
right that this is another ball game altogether (my "home-game" analogy
was an understatement -- and the metaphors are as mixed as nuts by
now...). So no analogies carry over. It's not that the rules have
changed. It's just that nothing remotely like this has ever come up
before. So, in particular, you are NOT entitled to help yourself to the
speed analogy in trying to refute Searle's Chinese Room Argument.
Because whereas it would have been no problem at all for Maxwell to
"bite the bullet" and claim that very slow EM oscillation was still
light, only it wasn't visible, one CANNOT say that very slow
symbol-manipulation is still thinking only it's... what?
"Unthinkable?" You took the words out of my mouth.

(4) Your point about the immunity of parallel processes to the Chinese
Room Argument (unlike similar points about speed, capacity or
complexity) has somewhat more prima facie force because it really is
based on something Searle can't take over all by himself, the way he
could with symbol manipulation. On the face of it, Searle couldn't BE
the parallel system that was passing the TT in Chinese in the same way
he could BE the serial symbol system, so he could not take the next
step and show that he would not be understanding Chinese if he were
(and hence that neither would the system he was duplicating).

This is why I suggested to Searle that his "Chinese Gym" Argument fails
to have the force of his original Chinese Room Argument, and is indeed
vulnerable to a "systems reply." It's also why I suggested the
"three-room" argument to Searle, which is completely in the
spirit of the original Chinese Room Argument and puts the burden of
evidence or argument on the essential parallelist, where it belongs.
Here is the critical excerpt from my comments on an earlier draft by
Searle:

> So I respectfully recommend that you jettison the Chinese Gym Argument
> and instead deal with connectionism by turning the Chinese Room
> Argument on its head, as follows. Suppose there are three rooms:
> 
> (1) In one there is a real Net (implemented as physical units, with
> real physical links, real excitatory/inhibitory interconnections
> real parallel distributed processing, real backpropping, etc.) that
> could pass the Turing Test in Chinese (Chinese symbols in, Chinese
> symbols out).
> 
> (2) In the second there is a computer simulation of (1) that likewise
> passes the TT in Chinese.
> 
> (3) In the third is Searle, performing ALL the functions of (2),
> likewise passing the Chinese TT (while still not understanding, of
> course).
> 
> Now the connectionists have only two choices:
> 
> Either they must claim that all three understand Chinese (in which case
> they are back up against the old Chinese Room Argument), or the
> essentialists among them will have to claim that (1) understands but (2)
> and (3) do not -- but without being able to give any functional reason 
> whatsoever why.

So this is what parallelism is up against. I also went on to query the
Connectionists on this, as follows (and received multiple replies, most
along the lines of the 1st two, which I include below):

> From:    Stevan Harnad
> To:      connectionists@cs.cmu.edu
> Subject: Parallelism, Real vs. Simulated: A Query
> 
>   "I have a simple question: What capabilities of PDP systems do and
>   do not depend on the net's actually being implemented in parallel,
>   rather than just being serially simulated? Is it only speed and
>   capacity parameters, or something more?"
> ----------------------------------------------------------------
> 
> (1) From: skrzypek@CS.UCLA.EDU (Dr. Josef Skrzypek)
> Cc: connectionists@cs.cmu.edu
> 
> Good (and dangerous) question. Applicable to Neural Nets in general
> and not only to PDP.
> 
> It appears that you can simulate anything that you wish. In principle
> you trade computation in space for computation in time. If you can
> make your time-slices small enough and complete all of the necessary
> computation within each slice there seem to be no reason to have
> neural networks. In reality, simulation of synchronized, temporal
> events taking place in a 3D network that allows for feedback pathways
> is rather cumbersome.
> 
> (2) From: Michael Witbrock <mjw@cs.cmu.edu>
> 
> I believe that none of the properties depend on parallel implementation.
> 
> There is a proof of the formal equivalence of continuous and discrete
> finite state automata, which I believe could be transformed to prove the
> formal equivalence of parallel and serially simulated pdp models.

Except for some equivocal stuff on "asynchronous" vs "synchronous"
processes, about which some claimed one thing and others claimed the
opposite, most respondents agreed that the parallel and serial
implementations were equivalent. Hence it is not true, as you write,
that parallel systems are "not threatened by [Searle's] Chinese Room
argument." They are, although someone may still come up with a
plausible reason why, although the computational difference is
nonexistent, the implementational difference is an essential one.

And that may, logically speaking, turn out to be (one of the)
answer(s) to the question of which of the "causal powers" of the brain
are actually relevant (and necessary/sufficient) for producing a mind.
I think Searle's Argument (and my Symbol Grounding Problem) have
effectively put pure symbol manipulation out of contention. I don't
think "the same, only faster, bigger, or more complex" holds much hope
either. And parallelism stands a chance only if someone can show what
it is about its implementation in that form, rather than in fast serial
symbolic form, is critical. My own favored candidate for the "relevant"
property, however, namely, sensory grounding, and sensory transduction
in particular, has the virtue of not only being, like parallelism,
invulnerable to the Chinese Argument (as I showed in "Minds, Machines
and Searle"), but also being a natural candidate for a solution to the
Symbol Grounding Problem, thereby, unlike paralellism, wearing the
reason WHY it's critical on its sleeve, so to speak.

(5) Finally, you write "We, and Searle, reject the Turing Test as a
sufficient condition for conscious intelligence." In this I must disagree
with both (or rather, all three) of you: The logic goes like this. So
far, only pure symbol crunching has been disqualified as a candidate
for being the sufficient condition for having a mind. But don't forget
that it was only a conjecture (and in my mind always a counterfactual
one) that the standard (language-only) Turing Test (only symbols in,
and symbols out), the TT, could be successfully passed by a pure symbol
cruncher. Searle's argument shows that IF the TT could be passed by
symbol crunching alone, THEN, because of the Chinese Room Argument, it
would not have a mind, and hence the TT is to be rejected.

Another possibility remains, however, which is that it is impossible to
successfully pass the TT with symbol crunching alone. The truth may
instead be that any candidate that could pass the TT would already have
to have and draw upon the causal power to pass the TTT, the Total
Turing Test, which includes all of our robotic, sensorimotor capacities
in the real world of objects. Now the TTT necessarily depends on
transduction, which is naturally and transparently immune to Searle's
Chinese Room Argument. Hence there is no reason to reject the TTT
(indeed, I would argue, there's no alternative to the TTT, which,
perhaps expanded to include neural function -- the "TTTT"? -- is simply
equivalent to empiricism!). And if a necessary condition for passing the
TT is the causal power to pass the TTT, then there's really no reason
left for rejecting the TT either.

Stevan Harnad

References:

Harnad, S. (1989) Minds, Machines and Searle. Journal of Experimental
and Theoretical Artificial Intelligence 1:

Harnad, S. (1990) The Symbol Grounding Problem. Physica D (In Press)

Preprints are available by email.
-- 
Stevan Harnad  Department of Psychology  Princeton University
harnad@confidence.princeton.edu       srh@flash.bellcore.com
harnad@elbereth.rutgers.edu    harnad@pucc.bitnet    (609)-921-7771