From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.ecf!utgpu!cs.utexas.edu!usc!elroy.jpl.nasa.gov!decwrl!mcnc!aurs01!sheol!throopw Mon May 25 14:07:06 EDT 1992
Article 5838 of comp.ai.philosophy:
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.ecf!utgpu!cs.utexas.edu!usc!elroy.jpl.nasa.gov!decwrl!mcnc!aurs01!sheol!throopw
>From: throopw@sheol.UUCP (Wayne Throop)
Newsgroups: comp.ai.philosophy
Subject: Re: Grounding: Virtual vs. Real
Summary: what reasons for a humaniform boundary; just what is such a boundary
Message-ID: <4782@sheol.UUCP>
Date: 22 May 92 03:06:16 GMT
References: <1992May21.045844.2833@Princeton.EDU>
Lines: 237

I'll preface this by saying I appreciate Stevan's patience in
sticking with this discussion.  I think I'm understanding much
better what he is getting at, despite my not finding his position
terribly attractive.

-> harnad@shine.Princeton.EDU (Stevan Harnad)
-> (1) HUMAN TTT: The only grounding worth talking about is the grounding
-> of the symbols inside a robot that is capable of passing the Total
-> Turing Test (TTT), i.e., a robot capable of all the sensorimotor and
-> verbal interactions with the real world of objects, events and states
-> of affairs that we are capable of -- Turing-indistinguishable from us
-> in that respect, as a matter of fact.

Why is it "the only grounding worth talking about"?  
Also, what is meant by "we are capable of"?
( More detailed facets of these questions occur below. )

-> (2) DEDICATED SYSTEMS: So let's not talk any more about "grounded"
-> type-writers and "grounded" computers. Neither type-writers nor
-> computers can do what it takes to pass the TTT. They can't see things
-> and they can't manipulate them. 

There are many humans who can't see things.  There are many humans who
can't manipulate things.  There are many who can't do either.  Many of
these humans pass the day-to-day casual "turing testing" folks do.  I
see no motivation to be more strict with machines. 

-> (3) REAL TRANSDUCERS: Transduction is the transformation of energy
-> from one form into another. The specific kind of transduction that
-> is relevant to TTT robots is sensorimotor transduction: They must
-> be capable of receiving the "shadows" of objects cast on their sensory
-> surfaces -- and then taking it from there.

Why do computers fail this qualification?  There are really and
truely transducers in my keyboard.  They receive the "shadows" of my
fingers and take it from there.

-> (The question is begged if the assumption is made that transduction is
-> trivial, just a thin surface that immediately goes into symbols:

Who said decoding a keyboard (or handwriting, or voice or any of a
number of things computers can do) is trivial?  And if they said it,
why do they think so?

-> But the minute you take away a robot's transduction capacity, you've
-> removed it's grounding, because a robot without sensorimotor capacity
-> cannot pass the TTT (hence it does not HAVE TTT capacity).

OK.  A standard continuum argument here.  A robot arm with five
fingers, a wrist, elbow, and shoulder joint would presumably be enough
of a "sensorimotor capacity" in a robot.  The arm types on a typewriter.
What if it has only two fingers (a claw arm)?  What about only one finger
and a single swivel joint (similar to the mouth-held stylus used by
the handicaped to type)?  What about using that stylus to point to
a letterboard?  What about a mechanism that, instead of pointing to
a character on a letterboard, rotates a ball to present the correct
character for inspection (think of a selectric typeball)?  What about
a 24-pin printhead?  A screen with phosphors?  

All of these get from something inside to robot to characters evident to
the outside world, and it seems to me that the possible ways of doing
this form a continuum.  It also seems that the start of this continuum
is enough for Stevan to grant grounding, and the end of it is not.  But
why is it relevant that most humans do it with five fingers, a wrist,
and elbow and a shoulder joint? Some humans (as already pointed out) do
NOT do it this way, yet I doubt it is reasonable to call their typing
the production of only squiggles and squoggles. 

Or in other words, we already do not subject humans to this sort of
TTT... why should we do so for machines?

-> (4) "BRAIN IN A VAT": This is the point where people usually start
-> to think of brains in vats, but real brains in vats are not computers
-> with their trivial peripherals detached: They are mostly analog
-> re-projections of the sensory and motor surfaces. It must be left
-> COMPLETELY moot what would be left of the insides of a TTT-capable
-> robot if you removed its transducers. 

I understand this to say that it is the boundary between the entity and
the surround (or environment) that matters, not the exact nature of the
surround, or the internals of the entity.  This notion has a great deal
of appeal.  But why should we draw this boundary in the way Stevan suggests?

Or to put it another way: I propose that the original Turing boundary
(that is, vaguely, any old boundary that can be trivially transformed
into language) is a good one.  In what relevant way is the total
humaniform boundary better?

-> What DEFINITELY cannot be
-> presupposed (without begging the question) is that there would just be
-> a computer left in there -- 

This seems to ammount to the assertion that "what's in there" may not be
computable, or that it may not be possible to interpret the internal
processing as transforms on sentences of a formal system.  OK, I'm game
to agree that it MAY not be...  but only that it MAY not be.  Now, IF it
is true that, starting at the fully-capable-human boundary between
entity and environment it is impossible to pass the TTT by connecting
relatively trivial transducers to a pure symbol engine, THEN we'd have
something.  But no such conclusion has been adequately supported, near
as I can tell. 

On the other hand, if it IS possible to separate the work of passing the
TTT into a symbol crunching core and a series of relatively trivial
translations to fully-capable-human-boundary, I find it very strange to
say that the symbol crunching core, considered as an entity, is much less
thoughtful than the core-plus-transducers. 

-> (5) TTT CAPACITY. Note that the criterion has always been TTT capacity
-> -- nothing about its causal history, nothing about the nature of the
-> actual objects in the world. 

If I follow this correctly, the issue is to choose the boundary, and
consider the transactions across this boundary.  The source of the
external implementation is moot, as is the source of the internal
implementation.  The question becomes then, is it POSSIBLE that there is
a pure-symbolic core trivially close ("trivial" relative to the
complexity of the core) to this boundary inside a TTT-passing entity. 
(At least, I HOPE I've got this right now.)

One thing to consider is, given the capability of producing
symbol-crunching core capable of passing the TT, it seems to me that it
is a trivial amount of extra work to make an entity that passes the TTT. 
Is there some reason to think this is NOT so?

-> So perhaps I should say the human BODY (brain and all) is grounded, to
-> short-circuit further sci-fi fantasies. 

I hope it's agreed that the points I raise in this post are NOT sf, nor
even implausible.  A major point I'm trying to make is that we do not
consider humans who have a boundary different from most of us to be
ungrounded, even when it it quite different.  Why should machines be any
different in this respect?

-> (6) HOW/WHY DOES TTT POWER GROUND SYMBOLS? Because in an ordinary,
-> ungrounded symbol system, whether it is static, like a book, or
-> dynamic, like a computer, symbols are manipulated only syntactically
-> (based on their shapes), 

I disagree.  The symbols can be manipulated by arbitrary criteria
in a computer.  Presumably I'm missing something that is meant by
"based on their shapes" here, but even in the most trivial computers
the processing of (say) text encoded as ascii is certainly not based
on the shape imaged by the displayed pixels.

Perhaps "based on their shapes" means something like "based on the
allowable sentences and transforms of some formal system" or the like. 
But this goes nowhere towards showing that "what humans do" inside their
TTT boundary is not isomorphic to some formal system. 

Further, the mere implementing of a formal system as a real computer and
program means that the symbols of the formal system are causally
connected to the real world, and have a prefered interpretation.  It is
preferred that a one in that video ram means a bright pixel, and a zero
means a dark one.  It is prefered that a one in certain IO registers
means that a particular key has been depressed.  And so on.  True, the
shape of the boundary is an unusually low-bandwidth one, compared to the
usual human one, but I don't see why this is relevant in principle. 

The "symbols" of, say, the latin alphabet are hardly the fundamental
symbols of a computer running a program.  That's a much, much higher
level.  Now presumably Stevan is, indeed, talking about that higher
level of symbol.  But if so, then, at that high level, it is false to
say that computers treat them by trivial formal rules.  The formal rules
at that level are far, far from being trivial, and it is certainly not
clear what the implications of this are. 

-> But that semantics is ungrounded, because it depends on our
-> interpretation. Independent of the interpretation we project on
-> it, there are no "meanings" in a book or a computer, because books and
-> computers are not the kinds of things that anything means anything to.

Saying "books and computers are not the kinds of things that anything
means anything to" is simply begging the question. 

-> The TTT puts a second set of constraints on the symbols in a symbol
-> system, over and above the constraint that (i) they must be
-> interpretable, by us, as being ABOUT something (which is what makes
-> them a symbol system in the first place, rather than random gibberish).
-> That second constraint is that (ii) the system (now a robot) must be
-> able to pick out (discriminate, categorize, manipulate, identify and
-> describe) the objects that its symbols are interpretable as being about
-> in a way that is coherent with the interpretation and indistinguishable
-> from the way we do it:

Seems uncontroversial so far.

-> The meanings of its symbols are then grounded
-> directly in its robotic capacity rather than just parasitic on the
-> meanings we project unto them.

But I don't see how to get from the above (that is, interpretable by
humans, and organized by the computer in ways we humans do) to any
requirement for a humaniform physical interface.  The "in the ways we
humans do" (or the stronger "indistinguishable from the way we do it"
used above) may be what leads there.  But if "indistinguishable" is
taken THAT seriously, we can rule out just about any nonhuman.  

SOME distinguishing criteria are not to be used, surely.  For example, I
imagine it won't count as a difference if X-rays show that the torques
to the joints of the robot are produced by electric motors instead of
muscle contractions, as it writes or types.  I imagine that's an issue
internal to the entity, and moot.  Yet we can clearly distinguish the
"way the robot" is producing those symbols from the way humans do.  And
what makes X-radiation part of the interface or not part?  I imagine it
won't count if we don't smell any body odor on the robot, especially
that we don't detect heat and moisture in its breath as it speaks.  Yet
again, it is clearly a way to tell the difference in the way it is
treating language.  What makes a change in the composition of air in
respiration and speech part of the interface or not part?

So some distinctions are worth making, and some are not.  The question
is, *exactly* *which* are worth making, and why. 

-> Is it now clearer why in order to keep this grounding intact, we cannot
-> eliminate the transducers and the TTT capacity and return to just a
-> symbol system that will merely bear the weight of a systematic
-> interpretation (e.g., as a virtual robot and its virtual world)? 

That is exactly my point.  What's going into and coming out of the
computer I'm typing into IS NOT SYMBOLS.  It is tactile senses (keyboard
and mouse), and actual, real effectors that have a prefered physical
interpretation (pixels on the screen, the vibration of the bell device,
pins impacting on paper leaving ink smears, and so on and on).  I see
nothing different in principle between this and human senses (tactile,
auditory, etc) and capabilities (finger motions, sometimes leaving ink
smears on paper, expulsion of air over vibrating cords, and so on). 

The question is, are those human finger movements, air expulsions, and
so on, being produced by a process interpretable as a formal system? We
have no way of knowing that.  Yet.  

We know that a computer's activity is produced in a way that IS so
interpretable.  We don't know that the human's is not.  Still less
do we know that all possible TTT-passing entities' are not.
--
Wayne Throop  ...!mcnc!dg-rtp!sheol!throopw


