Turing

From Tom Bourbon [920914]

Penni Subin [920913] gave what I think is a very nice reply to Rick
Marken's [920910] post concerning the Turing Test. The example of
judging another person's sex (gender) is elegant.
   In nearly all of my recent presentations and demonstrations of
PCT tasks in which people, or hands, or models, or combinations of
those interact, I have opted to sodel interacting with a
person, to make the point that you can scarcely tell the difference
between that condition and two people interacting. That seems to
me to be precisely the sort of test-demonstration Turing had in mind
in his 1950 article. I have come to think of such demonstrations as
The Second Test (along with the test for the controlled variable) that
PCT can off as an alternative to "business as usual" in the behavioral,
life and cognitive sciences. AI people (at least some of them) do
still accept a Turing Test as evidence concerning a model or theory. That
gives us a great opening to show the wide range of phenomena in which the
PCT model passes the test.

penni sibun (920914.1300):

   [From Rick Marken (920914.1200)]

   >female. maybe the analogy will be helpful. how do we know if someone
   >is male or female?

   >if *sex* is decided in such a ``superficial'' way, then why not the
   >ability to think?

   This analogy rests on the assumption that behavior is to thinking as
   gender is to chromosomes.

no, it rests on the assumption that, like intelligence, sex is
something we make judgements about when we consider people or other
organisms.

  Apparently, you assume that the "superficial"
   signs of gender (malenes vs femaleness) are an output of chromosome
structure:

   xy --> maleness
   xx --> femaleness

i said no such thing. how could you possibly have got that from my
post? i listed several counterexamples to your ``outputs.'' i think
you come to things w/ such preconceptions that i barely understand
what you're talking about when you summarize what you think i said.
some of this is undoubtedly my fault; i'm hardly perfect at presenting
my ideas. but i find you especially difficult to communicate w/.

   So, the "sex text" looks at observed maleness or femaleness and guesses
   at the "real" sex of the person (their chromosome structure).

no. the sex test looks at a person and guesses whether they are
female or male. period. the sex test doesn't care about what's
inside. it only has access to what you consider ``superficial''
things.

   The above
   assumptions make it reasonable to think that you can "work backwards",
   inferring chromosome structure from appearance.

no. the test doesn't work backwards; it doesn't have anything to work
backwards to.

   I agree that this is just how Turing looked at his test for mind.

well, we're obviously not agreeing at all.

    He
   assumed that "mental behavior" was an output of mental structure, and
   "non-mental behavior" was what resulted when there was no mind:

   mind --> mental behavior
   no-mind --> non-mental behavior.

no. when you go on like this, i see so clearly preston's point that
some people are so locked into the behaviorist/mentalist mindframe
that they can't talk about anything else. (did you ever get her
paper?)

i think you're a smart and articulate person, btw. i just think
you're terrifically wedged in this area. it's not my job to unwedge
you; but maybe one day one of my stories will make you look at things
just a little differently.

cheers.

        --penni

[From Rick Marken (920914.1200)]

penni sibun (920913.2000) says (to me):

that's ok; i don't usually read you since i know what you're going to
say.

ouch!

however, you did pique my curiosity by mentioning the turing
test--even though i knew what you were going to say ;-}.

turing based his test on what was apparently a sort of parlor game in
which the interrogator has to decide which is male and which is
female. maybe the analogy will be helpful. how do we know if someone
is male or female?

if *sex* is decided in such a ``superficial'' way, then why not the
ability to think?

This analogy rests on the assumption that behavior is to thinking as
gender is to chromosomes. Apparently, you assume that the "superficial"
signs of gender (malenes vs femaleness) are an output of chromosome structure:

xy --> maleness
xx --> femaleness

So, the "sex text" looks at observed maleness or femaleness and guesses
at the "real" sex of the person (their chromosome structure). The above
assumptions make it reasonable to think that you can "work backwards",
inferring chromosome structure from appearance.

I agree that this is just how Turing looked at his test for mind. He
assumed that "mental behavior" was an output of mental structure, and
"non-mental behavior" was what resulted when there was no mind:

mind --> mental behavior
no-mind --> non-mental behavior.

So the Turing test is based on the same idea as the sex test; look
at observed behavior and work backwards to infer whether or not the
behavior was generated by a system with a mind. If you see "mental
behavior" then guess mind; if you see non-mental behavior then
guess "no mind".

The point of my post (if you are still reading) was precisely this:
the Turing test is Behavioristic in the sense that it assumes that
mental behavior is the OUTPUT of a mental system. This assumption
is also the basis of cognitive psychology. The idea is that
you can infer something about the ming by looking at behavior
because behavior is an OUTPUT of mind.

I brought this up because, according to PCT, the Turing test
(as Turing described it) cannot work; behavior (results that are
intentionally produced by the organism) is NOT OUTPUT -- it is
controlled INPUT (ie. perceptions. The PCT model of mind
produces a 3rd alternative to the two above, namely:

mind -->input< -- actions
      > "behavior" |

The mind (reference signals) specifies intended levels of input.
The system acts to keep these inputs at their intended levels (but
it does not cause these actions; these actions are caused by the
disturbance - resistant characteristics of the system).

An observer might call either the actions or the observed correlate
of the intended input the "behavior" of the system. In PCT, we reserve
the word "behavior" for the intended inputs; actions are not really
the behavior of the system since the system doesn't really determine what
those actions will be. But actions (or their side effects as seen by
the observer) are what are most likely to be seen as behavior. These
actions, however, don't really reveal what Turing is trying to get at --
whether or not he is dealing with a system with a mind.

What Turing really needs to do (if the PCT model of the system is
correct) is figure out whether or not he is dealing with a system
whose actions are aimed at protecting the state of an input
variable (a variable that is determined by a mind). To do this,
Turing would have to know that there might be a controlled variable
involved (the "input" in the diagram) and test to see if that variable
is being protected from disturbance. This process is called "the test
for the controlled variable". It differs from the Turing test only
inasmuch as it clearly specifies QUANTITATIVELY how a system with
a mind differs from a system without one; the system with a mind is
busy trying to control input variables (at least if it is the kind of
mental system assumed by PCT -- a purposeful mental system). The no-mind
system is not controlling an input.

You cannot tell whether or not a system is controlling an input by
simply watching to see if its behavior "looks" intelligent. But this
is what Turing suggested WAS possible. You "poke" at the system
(with questions) and watch to see how it responds. If the response if
judged "intelligent" then you guess that it was made by a mental system.
This approach is not only unnecessarily subjective (as you hint at in
your "sex" example) -- it is bound to give the wrong results when you
are dealing with a mental system that is organized as a closed loop
control system. The failure of the Turing test is demonstrated by
my "mind read" and "find mind" programs (since you don't read my
stuff I guess I can count on your not reading about the mind read
program in my Mind Readings book). The "find mind" program, which you
will be happy to learn is NOT written up anywhere, is a good
example of the Turing Test. One of five numbers is actually controlling
it's position but all five are "doing" pretty complex movements. I
could have each one trace out the form of a character punched into
the keyboard -- their movement in reponse to the input is like their
response to a Turing question. I could even have only one consistently
answer correctly. But, still, all but one are generating behavioral
output. Only one is actually controlling an input -- and this is the
only one with a "mind" -- a reference for it's position. This one
is easily detected by disturbing it's position (with the mouse).

I doubt that you are still there, penni, but for the benefit of anyone
who still is, the point is the the Turing Test (as it is described in
Turing's article -- yes, I read it) is based on a behaviorist conception
of behavior -- ie. that behavior is an output generated by mind events.
PCT says that this concept of behavior is demonstrably false. So there is
no way to distinguish mind generated from non-mind generated behavior by
just LOOKING at characteristics of the behavior itself. You CAN,
however, distinguish mind generated from non-mind generated behavior
using the Powers Test (the test for the controlled variable) which
involves:
1. identifying a possible controlled variable
2. produce disturbances to the variable (events that should change
the hypothetical controlled variable if it is NOT controlled)
3. monitor state of hypothetical controlled variable under continuous
disturbance
4. look for lack of effect of disturbance
5. if there is no effect, make sure system being tested can perceive
the variable (control stops when variable is obscured) and can
effect it

Turing was a smart, swell guy and all but, unfortunately, he didn't
understand control (who did at the time?) so his test (like everything
else in the life sciences) was based on the idea that behavior is OUTPUT.
An understandable (but in 1992 a quite unnecessary) misconception.

Best regards

Rick

···

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

Richard S. Marken USMail: 10459 Holman Ave
The Aerospace Corporation Los Angeles, CA 90024
E-mail: marken@aero.org
(310) 336-6214 (day)
(310) 474-0313 (evening)

(eric harnden: 920915.1230)

ok, i think i'm coming a little closer, here. mr marken's post on the
turing test contained a thumbnail of the pct perspective which makes it
a little easier for me to understand why his insistence that ai should
not focus on modeling behavioral outputs has been so strongly stated,
and am finding myself more in agreement. i think that part of my
resistance to this point has been a degree of confusion over the
importance of the specific terminology. frankly, i don't think it matters
where one puts the 'behavior' label. my own approach has always been
a cybernetic one, and the shifting of a label doesn't change any of my
(mental or experimental) model structures. i also think that such a
(perhaps misplaced?) emphasis on nomenclature may be partly responsible
for an earlier interaction between mr marken and ms sibun. given the
interaction between entity and environment as the basis for discussion,
the question of the locus of mind is largely philosophical. as i told
ms sibun in a private memo, i found some of the material that she
presented interesting in as far it related to ideas about mind that i have
picked up from bateson. but so what? or so what about objections to those
views? i model the mechanism, not the philosophy.
one way in which i am now more confused, however... mr marken said something
about the organism not being the cause of behavior. that behavior is
'caused' by environmental disturbance of the organism's perceptions. on the
strictly structural level i'm willing to accept this, given the local
definition of behavior. but, since the purpose of behavior is adjustment of
perceptions toward reference values, then that which is caused by environmental
disturbance must be the organism's actions... which is precisely the
'normal' definition of behavior. is there a contradiction here? sorry if
the answer to this question is made glaringly clear by closer reading
of some part of BCP which i have skimmed/forgotten. also, btw, note that
this question falls in the category just discussed. i don't really see how
the distinction affects the model.

-----------< Cognitive Dissonance is a 20th Century Art Form >-----------
Eric Harnden (Ronin)
<HARNDEN@AUVM.BITNET> or <HARNDEN@AMERICAN.EDU>
The American University Physics Dept.
4400 Mass. Ave. NW, Washington, DC, 20016-8058
(202) 885-2748
---------------------< Join the Cognitive Dissidents >-------------------

[From Jeff Hunter (920924 - 2)]

(Rick Marken 920910.0900)
But I do think that the Turing
Test is an EXCELLENT example of the behavioristic basis of AI (and
cognitive science) etc. Harnad's contribution makes it even clearer -- he
suggests a Total Turning Test meaning its not enough to get a simulation
to answer questions like a real person -- you must also get it to
behave in all ways like a real person -- ie -- brush teeth,
play soccer, build model airplanes, etc.

  Harnad's addition to the Turing test is spurious. Steven Hawking
(A Brief History of Time) cannot brush teeth, play soccer, or build
model airplanes. He only talks through a device very much like a
teletype. Yet he is generally considered intelligent.

[Rick Marken (920924.1000)]

Maybe not clear, but not misguided. Of course the Turing test can be
done as a test for the controlled variable; I've said that over and over
again. My point is that it is NOT done that way (nor is psychological
researh done that way).

  Well here's an excerpt from Turing (via Hofstadter).

Interrogator: In the first line of your sonnet which reads "Shall I
  compare thee to a summer's day", would not a "spring day" do
  as well or better?

Witness: It wouldn't scan.

Interrogator: How about a "winter's day"? That would scan all right.

Witness: Yes, but nobody wants to be compared to a winter's day.

  The interrogator is clearly trying to find which variables
in the sonnet the witness is trying to maintain against disturbance.
Turing probably would have phrased it as trying to discover what
internal concepts and goals were held by the witness. PCT talk
would describe these as CEV's and references.

... I would be thrilled to find evidence that someone, somewhere
had been doing the Turing Test as a test for controlled variables.

  A slight frission running down your spine? :slight_smile:

      ... Jeff

···

--
De apibus semper dubitandum est - Winni Ille Pu