harnad@mind.UUCP (Stevan Harnad) (09/27/86)
The following are the Summary and Abstract, respectively, of two papers
I've been giving for the past year on the colloquium circuit. The first
is a joint critique of Searle's argument AND of the symbolic approach
to mind-modelling, and the second is an alternative proposal and a
synthesis of the symbolic and nonsymbolic approach to the induction
and representation of categories.
I'm about to publish both papers, but on the off chance that
there is a still a conceivable objection that I have not yet rebutted,
I am inviting critical responses. The full preprints are available
from me on request (and I'm still giving the talks, in case anyone's
interested).
***********************************************************
Paper #1:
(Preprint available from author)
MINDS, MACHINES AND SEARLE
Stevan Harnad
Behavioral & Brain Sciences
20 Nassau Street
Princeton, NJ 08542
Summary and Conclusions:
Searle's provocative "Chinese Room Argument" attempted to
show that the goals of "Strong AI" are unrealizable.
Proponents of Strong AI are supposed to believe that (i) the
mind is a computer program, (ii) the brain is irrelevant,
and (iii) the Turing Test is decisive. Searle's point is
that since the programmed symbol-manipulating instructions
of a computer capable of passing the Turing Test for
understanding Chinese could always be performed instead by a
person who could not understand Chinese, the computer can
hardly be said to understand Chinese. Such "simulated"
understanding, Searle argues, is not the same as real
understanding, which can only be accomplished by something
that "duplicates" the "causal powers" of the brain. In the
present paper the following points have been made:
1. Simulation versus Implementation:
Searle fails to distinguish between the simulation of a
mechanism, which is only the formal testing of a theory, and
the implementation of a mechanism, which does duplicate
causal powers. Searle's "simulation" only simulates
simulation rather than implementation. It can no more be
expected to understand than a simulated airplane can be
expected to fly. Nevertheless, a successful simulation must
capture formally all the relevant functional properties of a
successful implementation.
2. Theory-Testing versus Turing-Testing:
Searle's argument conflates theory-testing and Turing-
Testing. Computer simulations formally encode and test
models for human perceptuomotor and cognitive performance
capacities; they are the medium in which the empirical and
theoretical work is done. The Turing Test is an informal and
open-ended test of whether or not people can discriminate
the performance of the implemented simulation from that of a
real human being. In a sense, we are Turing-Testing one
another all the time, in our everyday solutions to the
"other minds" problem.
3. The Convergence Argument:
Searle fails to take underdetermination into account. All
scientific theories are underdetermined by their data; i.e.,
the data are compatible with more than one theory. But as
the data domain grows, the degrees of freedom for
alternative (equiparametric) theories shrink. This
"convergence" constraint applies to AI's "toy" linguistic
and robotic models as well, as they approach the capacity to
pass the Total (asympototic) Turing Test. Toy models are not
modules.
4. Brain Modeling versus Mind Modeling:
Searle also fails to note that the brain itself can be
understood only through theoretical modeling, and that the
boundary between brain performance and body performance
becomes arbitrary as one converges on an asymptotic model of
total human performance capacity.
5. The Modularity Assumption:
Searle implicitly adopts a strong, untested "modularity"
assumption to the effect that certain functional parts of
human cognitive performance capacity (such as language) can
be be successfully modeled independently of the rest (such
as perceptuomotor or "robotic" capacity). This assumption
may be false for models approaching the power and generality
needed to pass the Total Turing Test.
6. The Teletype versus the Robot Turing Test:
Foundational issues in cognitive science depend critically
on the truth or falsity of such modularity assumptions. For
example, the "teletype" (linguistic) version of the Turing
Test could in principle (though not necessarily in practice)
be implemented by formal symbol-manipulation alone (symbols
in, symbols out), whereas the robot version necessarily
calls for full causal powers of interaction with the outside
world (seeing, doing AND linguistic understanding).
7. The Transducer/Effector Argument:
Prior "robot" replies to Searle have not been principled
ones. They have added on robotic requirements as an
arbitrary extra constraint. A principled
"transducer/effector" counterargument, however, can be based
on the logical fact that transduction is necessarily
nonsymbolic, drawing on analog and analog-to-digital
functions that can only be simulated, but not implemented,
symbolically.
8. Robotics and Causality:
Searle's argument hence fails logically for the robot
version of the Turing Test, for in simulating it he would
either have to USE its transducers and effectors (in which
case he would not be simulating all of its functions) or he
would have to BE its transducers and effectors, in which
case he would indeed be duplicating their causal powers (of
seeing and doing).
9. Symbolic Functionalism versus Robotic Functionalism:
If symbol-manipulation ("symbolic functionalism") cannot in
principle accomplish the functions of the transducer and
effector surfaces, then there is no reason why every
function in between has to be symbolic either. Nonsymbolic
function may be essential to implementing minds and may be a
crucial constituent of the functional substrate of mental
states ("robotic functionalism"): In order to work as
hypothesized, the functionalist's "brain-in-a-vat" may have
to be more than just an isolated symbolic "understanding"
module -- perhaps even hybrid analog/symbolic all the way
through, as the real brain is.
10. "Strong" versus "Weak" AI:
Finally, it is not at all clear that Searle's "Strong
AI"/"Weak AI" distinction captures all the possibilities, or
is even representative of the views of most cognitive
scientists.
Hence, most of Searle's argument turns out to rest on
unanswered questions about the modularity of language and
the scope of the symbolic approach to modeling cognition. If
the modularity assumption turns out to be false, then a
top-down symbol-manipulative approach to explaining the mind
may be completely misguided because its symbols (and their
interpretations) remain ungrounded -- not for Searle's
reasons (since Searle's argument shares the cognitive
modularity assumption with "Strong AI"), but because of the
transdsucer/effector argument (and its ramifications for the
kind of hybrid, bottom-up processing that may then turn out
to be optimal, or even essential, in between transducers and
effectors). What is undeniable is that a successful theory
of cognition will have to be computable (simulable), if not
exclusively computational (symbol-manipulative). Perhaps
this is what Searle means (or ought to mean) by "Weak AI."
*************************************************************
Paper #2:
(To appear in: "Categorical Perception"
S. Harnad, ed., Cambridge University Press 1987
Preprint available from author)
CATEGORY INDUCTION AND REPRESENTATION
Stevan Harnad
Behavioral & Brain Sciences
20 Nassau Street
Princeton NJ 08542
Categorization is a very basic cognitive activity. It is
involved in any task that calls for differential responding,
from operant discrimination to pattern recognition to naming
and describing objects and states-of-affairs. Explanations
of categorization range from nativist theories denying that
any nontrivial categories are acquired by learning to
inductivist theories claiming that most categories are learned.
"Categorical perception" (CP) is the name given to a
suggestive perceptual phenomenon that may serve as a useful
model for categorization in general: For certain perceptual
categories, within-category differences look much smaller
than between-category differences even when they are of the
same size physically. For example, in color perception,
differences between reds and differences between yellows
look much smaller than equal-sized differences that cross
the red/yellow boundary; the same is true of the phoneme
categories /ba/ and /da/. Indeed, the effect of the category
boundary is not merely quantitative, but qualitative.
There have been two theories to explain CP effects. The
"Whorf Hypothesis" explains color boundary effects by
proposing that language somehow determines our view of
reality. The "motor theory of speech perception" explains
phoneme boundary effects by attributing them to the patterns
of articulation required for pronunciation. Both theories
seem to raise more questions than they answer, for example:
(i) How general and pervasive are CP effects? Do they occur
in other modalities besides speech-sounds and color? (ii)
Are CP effects inborn or can they be generated by learning
(and if so, how)? (iii) How are categories internally
represented? How does this representation generate
successful categorization and the CP boundary effect?
Some of the answers to these questions will have to come
from ongoing research, but the existing data do suggest a
provisional model for category formation and category
representation. According to this model, CP provides our
basic or elementary categories. In acquiring a category we
learn to label or identify positive and negative instances
from a sample of confusable alternatives. Two kinds of
internal representation are built up in this learning by
"acquaintance": (1) an iconic representation that subserves
our similarity judgments and (2) an analog/digital feature-
filter that picks out the invariant information allowing us
to categorize the instances correctly. This second,
categorical representation is associated with the category
name. Category names then serve as the atomic symbols for a
third representational system, the (3) symbolic
representations that underlie language and that make it
possible for us to learn by "description."
This model provides no particular or general solution to the
problem of inductive learning, only a conceptual framework;
but it does have some substantive implications, for example,
(a) the "cognitive identity of (current) indiscriminables":
Categories and their representations can only be provisional
and approximate, relative to the alternatives encountered to
date, rather than "exact." There is also (b) no such thing
as an absolute "feature," only those features that are
invariant within a particular context of confusable
alternatives. Contrary to prevailing "prototype" views,
however, (c) such provisionally invariant features MUST
underlie successful categorization, and must be "sufficient"
(at least in the "satisficing" sense) to subserve reliable
performance with all-or-none, bounded categories, as in CP.
Finally, the model brings out some basic limitations of the
"symbol-manipulative" approach to modeling cognition,
showing how (d) symbol meanings must be functionally
anchored in nonsymbolic, "shape-preserving" representations
-- iconic and categorical ones. Otherwise, all symbol
interpretations are ungrounded and indeterminate. This
amounts to a principled call for a psychophysical (rather
than a neural) "bottom-up" approach to cognition.rush@cwrudg.UUCP (rush) (10/01/86)
In article <158@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: >6. The Teletype versus the Robot Turing Test: > >For example, the "teletype" (linguistic) version of the Turing... > whereas the robot version necessarily >calls for full causal powers of interaction with the outside >world (seeing, doing AND linguistic understanding). > Uh...I never heard of the "robot version" of the Turing Test, could someone please fill me in?? I think that understanding the reasons for such a test would help me (I make no claims for anyone else) make some sense out of the rest of this article. In light of my lack of knowledge, please forgive my presumption in the following comment. >7. The Transducer/Effector Argument: > >A principled >"transducer/effector" counterargument, however, can be based >on the logical fact that transduction is necessarily >nonsymbolic, drawing on analog and analog-to-digital >functions that can only be simulated, but not implemented, >symbolically. > [ I know I claimed no commentary, but it seems that this argument depends heavily on the meaning of the term "symbol". This could be a problem that only arises when one attempts to implement some of the stranger possibilities for symbolic entities. ] Richard Rush - Just another Jesus freak in computer science decvax!cwruecmp!cwrudg!rush
harnad@mind.UUCP (Stevan Harnad) (10/02/86)
In his commentary-not-reply to my <158@mind.UUCP>, Richard Rush <150@cwrudge.UUCP> asks: (1) > I never heard of the "robot version" of the Turing Test, > could someone please fill me in? He also asks (in connection with my "transducer/effector" argument) about the analog/symbolic distinction: (2) > I know I claimed no commentary, but it seems that this argument > depends heavily on the meaning of the term "symbol". This could > be a problem that only arises when one attempts to implement some > of the stranger possibilities for symbolic entities. In reply to (1): The linguistic version of the turing test (turing's original version) is restricted to linguistic interactions: Language-in/Language-out. The robotic version requires the candidate system to operate on objects in the world. In both cases the (turing) criterion is whether the system can PERFORM indistinguishably from a human being. (The original version was proposed largely so that your judgment would not be prejudiced by the system's nonhuman appearance.) On my argument the distinction between the two versions is critical, because the linguistic version can (in principle) be accomplished by nothing but symbols-in/symbols-out (and symbols in between) whereas the robotic version necessarily calls for non-symbolic processes (transducer, effector, analog and A/D). This may represent a substantive functional limitation on the symbol-manipulative approach to the modeling of mind (what Searle calls "Strong AI"). In reply to (2): I don't know what "some of the stranger possibilities for symbolic entities" are. I take symbol-manipulation to be syntactic: Symbols are arbitrary tokens manipulated in accordance with certain formal rules on the basis of their form rather than their meaning. That's symbolic computation, whether it's done by computer or by paper-and-pencil. The interpretations of the symbols (and indeed of the manipulations and their outcomes) are ours, and are not part of the computation. Informal and figurative meanings of "symbol" have little to do with this technical concept. Symbols as arbitrary syntactic tokens in a formal system can be contrasted with other kinds of objects. The ones I singled out in my papers were "icons" or analogs of physical objects, as they occur in the proximal physical input/output in transduction, as they occur in the A-side of A/D and D/A transformations, and as they may function in any part of a hybrid system to the extent that their functional role is not merely formal and syntactic (i.e., to the extent that their form is not arbitrary and dependent on convention and interpretation to link it to the objects they "stand for," but rather, the link is one of physical resemblance and causality). The category-representation paper proposes an architecture for such a hybrid system. Stevan Harnad princeton!mind!harnad
me@utai.UUCP (Daniel Simon) (10/06/86)
In article <160@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: > >In reply to (1): The linguistic version of the turing test (turing's >original version) is restricted to linguistic interactions: >Language-in/Language-out. The robotic version requires the candidate >system to operate on objects in the world. In both cases the (turing) >criterion is whether the system can PERFORM indistinguishably from a human >being. (The original version was proposed largely so that your >judgment would not be prejudiced by the system's nonhuman appearance.) > I have no idea if this is a relevant issue or a relevant place to bring it up, but this whole business of the Turing test makes me profoundly suspicious. For example, we all know about Weizenbaum's ELIZA, which, he claimed, convinced many clever, relatively computer-literate (for their day) people that it was intelligent. This fact leads me to some questions which, in my view, ought to be seriously addressed before the phrase "Turing test" is bandied about (and probably already have been addressed, but I didn't notice, and will thank everybody in advance for telling me where to find a treatment of them and asking me to kindly buzz off): 1) To what extent is our discernment of intelligent behaviour context- dependent? ELIZA was able to appear intelligent because of the clever choice of context (in a Rogerian therapy session, the kind of dull, repetitive comments made by ELIZA seem perfectly appropriate, and hence, intelligent). Mr. Harnad has brought up the problem of physical appearance as a prejudicing factor in the assessment of "human" qualities like intelligence. Might not the robot version lead to the opposite problem of testers being insufficiently skeptical of a machine with human appearance (or even of a machine so unlike a human being in appearance that mildly human-like behaviour takes on an exaggerated significance in the tester's mind)? Is it ever possible to trust the results of any instance of the test as being a true indicator of the properties of the tested entity itself, rather than those of the environment in which it was tested? 2) Assuming that some "neutral" context can be found which would not "distort" the results of the test (and I'm not at all convinced that such a context exists, or even that the idea of such a context has any meaning), what would be so magic about the level of perceptiveness of the shrewdest, most perspicacious tester available, that would make his inability to distinguish man from machine in some instance the official criterion by which to judge intelligence? In short, what does passing (or failing) the Turing test really mean? 3) If the Turing test is in fact an unacceptable standard, and building a machine that can pass it an inappropriate goal (and, as questions 1 and 2 have probably already suggested, this is what I strongly suspect), are there more appropriate means by which we could evaluate the human-like or intelligent properties of an AI system? In effect, is it possible to formulate the qualities that constitute intelligence in a manner which is more intuitively satisfying than the standard AI stuff about reasoning, but still more rigorous than the Turing test? As I said, I don't know if my questions are legitimate, or if they have already been satisfactorily resolved, or if they belong elsewhere; I merely bring them up here because this is the first place I have seen the Turing test brought up in a long time. I am eager to see what others have to say on the subject. >Stevan Harnad >princeton!mind!harnad Daniel R. Simon "Look at them yo-yo's, that's the way to do it Ya go to grad school, get your PhD"
drew@ukma.uky.csnet (Andrew Lawson) (10/09/86)
In article <160@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: > >On my argument the distinction between the two versions is critical, >because the linguistic version can (in principle) be accomplished by >nothing but symbols-in/symbols-out (and symbols in between) whereas >the robotic version necessarily calls for non-symbolic processes >(transducer, effector, analog and A/D). This is not clear. When I look at my surroundings, you are no more than a symbol (just as is anything outside of my being). Remember that "symbol" is not rigidly defined most of the time. When I recognize the symbol of a car heading toward me, I respond by moving out of the way. This is not essentially different from a linguistic system recognizing a symbol and responding with another symbol. -- Drew Lawson cbosgd!ukma!drew "Parts is parts." drew@uky.csnet drew@UKMA.BITNET
harnad@mind.UUCP (Stevan Harnad) (10/10/86)
In response to what I wrote in article <160@mind.UUCP>, namely: >On my argument the distinction between the two versions >[of the turing test] is critical, >because the linguistic version can (in principle) be accomplished by >nothing but symbols-in/symbols-out (and symbols in between) whereas >the robotic version necessarily calls for non-symbolic processes >(transducer, effector, analog and A/D). Drew Lawson replies: > This is not clear. When I look at my surroundings, you are no > more than a symbol (just as is anything outside of my being). > Remember that "symbol" is not rigidly defined most of the time. > When I recognize the symbol of a car heading toward me, I respond > by moving out of the way. This is not essentially different from > a linguistic system recognizing a symbol and responding with another > symbol. It's important, when talking about what is and is not a symbol, to speak literally and not symbolically. What I mean by a symbol is an arbitrary formal token, physically instantiated in some way (e.g., as a mark on a piece of paper or the state of a 0/1 circuit in a machine) and manipulated according to certain formal rules. The critical thing is that the rules are syntactic, that is, the symbol is manipulated on the basis of its shape only -- which is arbitrary, apart from the role it plays in the formal conventions of the syntax in question. The symbol is not manipulated in virtue of its "meaning." Its meaning is simply an interpretation we attach to the formal goings-on. Nor is it manipulated in virtue of a relation of resemblance to whatever "objects" it may stand for in the outside world, or in virtue of any causal connection with them. Those relations are likewise mediated only by our interpretations. This is why the distinction between symbolic and nonsymbolic processes in cognition (and robotics) is so important. It will not do to simply wax figurative on what counts as a symbol. If I'm allowed to use the word metaphorically, of course everything's a "symbol." But if I stick to a specific, physically realizable sense of the word, then it becomes a profound theoretical problem just exactly how I (or any device) can recognize you, or a car, or anything else, and how I (or it) can interact with such external objects robotically. And the burden of my paper is to show that this capacity depends crucially on nonsymbolic processes. Finally, apart from the temptation to lapse into metaphor about "symbols," there is also the everpresent lure of phenomenology in contemplating such matters. For, apart from my robotic capacity to interact with objects in the world -- to recognize them, manipulate them, name them, describe them -- there is also my concsiousness: My subjective sense, accompanying all these capacities, of what it's like (qualitatively) to recognize, manipulate, etc. That, as I argue in another paper (and only hint at in the two under discussion), is a problem that we'd do best to steer clear of in AI, robotics and cognitive modeling, at least for the time being. We already have our hands full coming up with a model that can successfully pass the (robotic and/or linguistic) turing test -- i.e., perform exactly AS IF it had subjective experiences, the way we do, while it successfully accomplishes all those clever things. Until we manage that, let's not worry too much about whether the outcome will indeed be merely "as if." Overinterpreting our tools phenomenologically is just as unproductive as overinterpreting them metaphorically. Stevan Harnad princeton!mind!harnad
harnad@mind.UUCP (Stevan Harnad) (10/10/86)
In response to my article <160@mind.UUCP>, Daniel R. Simon asks: > 1) To what extent is our discernment of intelligent behaviour > context-dependent?...Might not the robot version [of the > turing test] lead to the...problem of testers being > insufficiently skeptical of a machine with human appearance? > ...Is it ever possible to trust the results of any > instance of the test...? My reply to these questions is quite explicit in the papers in question: The turing test has two components, (i) a formal, empirical one, and (ii) an informal, intuitive one. The formal empirical component (i) is the requirement that the system being tested be able to generate human performance (be it robotic or linguistic). That's the nontrivial burden that will occupy theorists for at least decades to come, as we converge on (what I've called) the "total" turing test -- a model that exhibits all of our robotic and lingistic capacities. The informal, intuitive component (ii) is that the system in question must perform in a way that is indistinguishable from the performance of a person, as judged by a person. It is not always clear which of the two components a sceptic is worrying about. It's usually (ii), because who can quarrel with the principle that a veridical model should have all of our performance capacities? Now the only reply I have for the sceptic about (ii) is that he should remember that he has nothing MORE than that to go on in the case of any other mind than his own. In other words, there is no rational reason for being more sceptical about robots' minds (if we can't tell their performance apart from that of people) than about (other) peoples' minds. The turing test is ALREADY the informal way we contend with the "other-minds" problem [i.e., how can you be sure anyone else but you has a mind, rather than merely acting AS IF it had a mind?], so why should we demand more in the case of robots? It's surely not because of any intuitive or a priori knowledge we have about the FUNCTIONAL basis of our own minds, otherwise we could have put those intuitive ideas to work in designing successful candidates for the turing test long ago. So, since we have absolutely no intuitive idea about the functional (symbolic, nonsymbolic, physical, causal) basis of the mind, our only nonarbitrary basis for discriminating robots from people remains their performance. As to "context," as I argue in the paper, the only one that is ultimately defensible is the "total" turing test, since there is no evidence at all that either capacities or contexts are modular. The degrees of freedom of a successful total-turing model are then reduced to the usual underdetermination of scientific theory by data. (It's always possible to carp at a physicist that his theoretic model of the universe "is turing-indistinguishable from the real one, but how can you be sure it's `really true' of the world?") > 2) Assuming that some "neutral" context can be found... > what does passing (or failing) the Turing test really mean? It means you've successfully modelled the objective observables under investigation. No empirical science can offer more. And the only "neutral" context is the total turing test (which, like all inductive contexts, always has an open end, namely, the everpresent possibility that things could turn out differently tomorrow -- philosophers call this "inductive risk," and all empirical inquiry is vulnerable to it). > 3) ...are there more appropriate means by which we > could evaluate the human-like or intelligent properties of an AI > system? ...is it possible to formulate the qualities that > constitute intelligence in a manner which is more intuitively > satisfying than the standard AI stuff about reasoning, but still > more rigorous than the Turing test? I don't think there's anything more rigorous than the total turing test since, when formulated in the suitably generalized way I describe, it can be seen to be identical to the empirical criterion for all of the objective sciences. Residual doubts about it come from four sources, as far as I can make out, and only one of these is legitimate. The legitimate one (a) is doubts about autonomous symbolic processes (that's what my papers are about). The three illegitimate ones (in my view) are (b) misplaced doubts about underdetermination and inductive risk, (c) misplaced hold-outs for the nervous system, and (d) misplaced hold-outs for consciousness. For (a), read my papers. I've sketched an answer to (b) above. The quick answer to (c) [brain bias] -- apart from the usual structure/function and multiple-realizability arguments in engineering, computer science and biology -- is that as one approaches the asymptotic Total Turing Test, any objective aspect of brain "performance" that anyone believes is relevant -- reaction time, effects of damage, effects of chemicals -- is legitimate performance data too, including microperformance (like pupillary dilation, heart-rate and perhaps even synactic transmission). I believe that sorting out how much of that is really relevant will only amount to the fine-tuning -- the final leg of our trek to theoretic Utopia, with most of the substantive theoretical work already behind us. Finally, my reply to (d) [mind bias] is that holding out for consciousness is a red herring. Either our functional attempts to model performance will indeed "capture" consciousness at some point, or they won't. If we do capture it, the only ones that will ever know for sure that we've succeeded are our robots. If we don't capture it, then we're stuck with a second level of underdetermination -- call it "subjective" underdetermination -- to add to our familiar objective underdetermination (b): Objective underdetermination is the usual underdetermination of objective theories by objective data; i.e., there may be more than one way to skin a cat; we may not happen to have converged on nature's way in any of our theories, and we'll never be able to know for sure. The subjective twist on this is that, apart from this unresolvable uncertainty about whether or not the objective models that fit all of our objective (i.e., intersubjective) observations capture the unobservable basis of everything that is objectively observable, there may be a further unresolvable uncertainty about whether or not they capture the unobservable basis of everything (or anything) that is subjectively observable. AI, robotics and cognitive modeling would do better to learn to live with this uncertainty and put it in context, rather than holding out for the un-do-able, while there's plenty of the do-able to be done. Stevan Harnad princeton!mind!harnad
cda@entropy.berkeley.edu (10/13/86)
In article <167@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes:
<as one approaches the
<asymptotic Total Turing Test, any objective aspect of brain
<"performance" that anyone believes is relevant -- reaction time,
<effects of damage, effects of chemicals -- is legitimate performance
<data too, including microperformance (like pupillary dilation,
<heart-rate and perhaps even synactic transmission).
Does this mean that in order to successfully pass the Total Turing Test,
a robot will have to be able to get high on drugs? Does this imply that the
ability of the brain to respond to drugs is an integral component of
intelligence? What will Ron, Nancy, and the DOD think of this idea?
Turing said that the way to give a robot free will was to incorporate
sufficient randomness into its actions, which I'm sure the DOD won't like
either.
It seems that intelligence is not exactly the quality our government is
trying to achieve in its AI hard and software.michaelm@bcsaic.UUCP (michael maxwell) (10/14/86)
In article <167@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: >...since there is no >evidence at all that either capacities or contexts are modular. Maybe I'm reading this out of context (not having read your books or papers), but could you explain this statement? I know of lots of evidence for the modularity of various aspects of linguistic behavior. In fact, we have a parser + grammar of English here that captures a large portion of English syntax, but has absolutely no semantics (yet). That is, it could parse Jabberwocky or your article (well, I can't quite claim that it would parse *all* of either one!) without having the least idea that your article is meaningful whereas Jabberwocky isn't (apart from an explanation by Humpty Dumpty). On the other hand, it wouldn't parse something like "book the table on see I", despite the fact that we might make sense of the latter (because of our world knowledge). Likewise, human aphasics often show similar deficits in one or another area of their speech or language understanding. If this isn't modular, what is? But as I say, maybe I don't understand what you mean by modular... -- Mike Maxwell Boeing Advanced Technology Center ...uw-beaver!uw-june!bcsaic!michaelm
franka@mmintl.UUCP (Frank Adams) (10/15/86)
In article <166@mind.UUCP> harnad@mind.UUCP writes: >What I mean by a symbol is an >arbitrary formal token, physically instantiated in some way (e.g., as >a mark on a piece of paper or the state of a 0/1 circuit in a >machine) and manipulated according to certain formal rules. The >critical thing is that the rules are syntactic, that is, the symbol is >manipulated on the basis of its shape only -- which is arbitrary, >apart from the role it plays in the formal conventions of the syntax >in question. The symbol is not manipulated in virtue of its "meaning." >Its meaning is simply an interpretation we attach to the formal >goings-on. Nor is it manipulated in virtue of a relation of >resemblance to whatever "objects" it may stand for in the outside >world, or in virtue of any causal connection with them. Those >relations are likewise mediated only by our interpretations. I see two problems with respect to this viewpoint. One is that relating purely symbolic functions to external events is essentially a solved problem. Digital audio recording, for example, works quite well. Robotic operations generally fail, when they do, not because of any problems with the digital control of an analog process, but because the purely symbolic portion of the process is inadequate. In other words, there is every reason to expect that a computer program able to pass the Turing test could be extended to one able to pass the robotic version of the Turing test, requiring additional development effort which is tiny by comparison (though likely still measured in man-years). Secondly, even in a purely formal environment, there turn out to be a lot of real things to talk about. Primitive concepts of time (before and after) are understandable. One can talk about nouns and verbs, sentences and conversations, self and other. I don't see any fundamental difference between the ability to deal with symbols as real objects, and the ability to deal with other kinds of real objects. Frank Adams ihnp4!philabs!pwa-b!mmintl!franka Multimate International 52 Oakland Ave North E. Hartford, CT 06108
me@utai.UUCP (Daniel Simon) (10/16/86)
In article <167@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: > >In response to my article <160@mind.UUCP>, Daniel R. Simon asks: > >> 1) To what extent is our discernment of intelligent behaviour >> context-dependent?...Might not the robot version [of the >> turing test] lead to the...problem of testers being >> insufficiently skeptical of a machine with human appearance? >> ...Is it ever possible to trust the results of any >> instance of the test...? > >My reply to these questions is quite explicit in the papers in >question: >The turing test has two components, (i) a formal, empirical one, >and (ii) an informal, intuitive one. The formal empirical component (i) >is the requirement that the system being tested be able to generate human >performance (be it robotic or linguistic). That's the nontrivial >burden that will occupy theorists for at least decades to come, as we >converge on (what I've called) the "total" turing test -- a model that >exhibits all of our robotic and lingistic capacities. By "nontrivial burden", do you mean the task of defining objective criteria by which to characterize "human performance"? If so, you are after the same thing as I am, but I fail to see what this has to do with the Turing test as originally conceived, which involved measuring up AI systems against observers' impressions, rather than against objective standards. Apparently, you're not really defending the Turing test at all, but rather something quite different. Moreover, you haven't said anything concrete about what this test might look like. On what foundation could such a set of defining characteristics for "human performance" be based? Would it define those attributes common to all human beings? Most human beings? At least one human being? How would we decide by what criteria to include observable attributes in our set of "human" ones? How could such attributes be described? Is such a set of descriptions even feasible? If not, doesn't it call into question the validity of seeking to model what cannot be objectively characterized? And if such a set of describable attributes is feasible, isn't it an indispensable prerequisite for the building of a working Turing-test-passing model? Please forgive my impertinent questions, but I haven't read your articles, and I'm not exactly clear about what this "total" Turing test entails. >The informal, >intuitive component (ii) is that the system in question must perform in a >way that is indistinguishable from the performance of a person, as >judged by a person. > >Now the only reply I have for the sceptic about (ii) is >that he should remember that he has nothing MORE than that to go on in >the case of any other mind than his own. In other words, there is no >rational reason for being more sceptical about robots' minds (if we >can't tell their performance apart from that of people) than about >(other) peoples' minds. The turing test is ALREADY the informal way we >contend with the "other-minds" problem [i.e., how can you be sure >anyone else but you has a mind, rather than merely acting AS IF it had >a mind?], so why should we demand more in the case of robots? ... > I'm afraid I must disagree. I believe that people in general dodge the "other minds" problem simply by accepting as a convention that human beings are by definition intelligent. For example, we use terms such as "autistic", "catatonic", and even "sleeping" to describe people whose behaviour would in most cases almost certainly be described as unintelligent if exhibited by a robot. Such people are never described as "unintelligent" in the sense of the word that we would use to describe a robot who showed the exact same behaviour patterns. Rather, we imply by using these terms that the people being described are human, and therefore *would* be behaving intelligently, but for (insert neurophysiological/psychological explanation here). This implicit axiomatic attribution of intelligence to humans helps us to avoid not only the "other minds" problem, but also the problem of assessing intelligence despite the effect of what I previously referred to loosely as the "context" of our observations. In short, we do not really use the Turing test on each other, because we are all well acquainted with how easily we can be fooled by contextual traps. Instead, we automatically associate intelligence with human beings, thereby making our intuitive judgment even less useful to the AI researcher working with computers or robots. >As to "context," as I argue in the paper, the only one that is >ultimately defensible is the "total" turing test, since there is no >evidence at all that either capacities or contexts are modular. The >degrees of freedom of a successful total-turing model are then reduced >to the usual underdetermination of scientific theory by data. (It's always >possible to carp at a physicist that his theoretic model of the >universe "is turing-indistinguishable from the real one, but how can >you be sure it's `really true' of the world?") > Wait a minute--You're back to component (i). What you seem to be saying is that the informal component (component (ii)) has no validity at all apart from the "context" of having passed component (i). The obvious conclusion is that component (ii) is superfluous; any system that passes the "total Turing test" exhibits "human behaviour", and hence must by definition be indistinguishable from a human to another human. >> 2) Assuming that some "neutral" context can be found... >> what does passing (or failing) the Turing test really mean? > >It means you've successfully modelled the objective observables under >investigation. No empirical science can offer more. And the only >"neutral" context is the total turing test (which, like all inductive >contexts, always has an open end, namely, the everpresent possibility >that things could turn out differently tomorrow -- philosophers call >this "inductive risk," and all empirical inquiry is vulnerable to it). > Again, you have all but admitted that the "total" Turing test you have described has nothing to do with the Turing test at all--it is a set of "objective observables" which can be verified through scientific examination. The thoughtful examiner and "comparison human" have been replaced with controlled scientific experiments and quantifiable results. What kinds of experiments? What kinds of results? WHAT DOES THE "TOTAL TURING TEST" LOOK LIKE? >> 3) ...are there more appropriate means by which we >> could evaluate the human-like or intelligent properties of an AI >> system? ...is it possible to formulate the qualities that >> constitute intelligence in a manner which is more intuitively >> satisfying than the standard AI stuff about reasoning, but still >> more rigorous than the Turing test? > >I don't think there's anything more rigorous than the total turing >test since, when formulated in the suitably generalized way I >describe, it can be seen to be identical to the empirical criterion for >all of the objective sciences... One question you haven't addressed is the relationship between intelligence and "human performance". Are the two synonymous? If so, why bother to make artificial humans when making natural ones is so much easier (not to mention more fun)? And if not, how does your "total Turing test" relate to the discernment of intelligence, as opposed to human-like behaviour? I know, I know. I ask a lot of questions. Call me nosy. > > >Stevan Harnad >princeton!mind!harnad Daniel R. Simon "We gotta install database systems Custom software delivery We gotta move them accounting programs We gotta port them all to PC's...."
harnad@mind.UUCP (Stevan Harnad) (10/16/86)
In reply to a prior iteration D. Simon writes: > I fail to see what [your "Total Turing Test"] has to do with > the Turing test as originally conceived, which involved measuring > up AI systems against observers' impressions, rather than against > objective standards... Moreover, you haven't said anything concrete > about what this test might look like. How about this for a first approximation: We already know, roughly speaking, what human beings are able to "do" -- their total cognitive performance capacity: They can recognize, manipulate, sort, identify and describe the objects in their environment and they can respond and reply appropriately to descriptions. Get a robot to do that. When you think he can do everything you know people can do formally, see whether people can tell him apart from people informally. > I believe that people in general dodge the "other minds" problem > simply by accepting as a convention that human beings are by > definition intelligent. That's an artful dodge indeed. And do you think animals also accept such conventions about one another? Philosophers, at least, seem to have noticed that there's a bit of a problem there. Looking human certainly gives us the prima facie benefit of the doubt in many cases, but so far nature has spared us having to contend with any really artful imposters. Wait till the robots begin giving our lax informal turing-testing a run for its money. > What you seem to be saying is that [what you call] > the informal component [(i) of the turing test -- > i. e., indistinguishability from a person, as judged by a > person] has no validity at all apart from the "context" of > having passed [your] component (i) [i.e., the generation of > our total cognitive performance capacity]. The obvious > conclusion is that component (ii) is superfluous. It's no more superfluous than, say, the equivalent component in the design of an artificial music composer. First you get it to perform in accordance with what you believe to be the formal rules of (diatonic) composition. Then, when it successfully performs according to the rules, see whether people like its stuff. Peoples' judgments, after all, were not only the source of those rules in the first place, but without the informal aesthetic sense that guided them, the rules would amount to just that -- meaningless acoustic syntax. Perhaps another way of putting it is that I doubt that what guides our informal judgments (and underlies our capacities) can be completely formalized in advance. The road to Total-Turing Utopia will probably be a long series of feedback cycles between the formal and informal components of the test before we ever achieve our final passing grade. > One question you haven't addressed is the relationship between > intelligence and "human performance". Are the two synonymous? > If so, why bother to make artificial humans... And if not, how > does your "total Turing test" relate to the discernment of > intelligence, as opposed to human-like behaviour? Intelligence is what generates human performance. We make artificial humans to implement and test our theories about the substrate of human performance capacity. And there's no objective difference between human and (turing-indistinguishably) human-like. > WHAT DOES THE "TOTAL TURING TEST" LOOK LIKE?... Please > forgive my impertinent questions, but I haven't read your > articles, and I'm not exactly clear about what this "total" > Turing test entails. Try reading the articles. ****** I will close with an afterthought on "blind" vs. "nonblind" turing testing that I had after the last iteration: In the informal component of the total turing test it may be arguable that a sceptic would give a robot a better run for its money if he were pre-alerted to the possibility that it was a robot (i.e., if the test were conducted "nonblind" rather than "blind"). That way the robot wouldn't be inheriting so much of the a priori benefit of the doubt that had accrued from our lifetime of successful turing-testing of biological persons of similar appearance (in our everyday informal solutions to the "other-minds" problem). The blind/nonblind issue does not seem critical though, since obviously the turing test is an open-ended one (and probably also, like all other empirical conjectures, confirmable only as a matter of degree); so we probably wouldn't want to make up our minds too hastily in any case. I would say that several years of having lived amongst us, as in the sci-fi movies, without arousing any suspicions -- and eliciting only shocked incredulity from its close friends once the truth about its roots was revealed -- would count as a pretty good outcome on a "blind" total turing test. Stevan Harnad princeton!mind!harnad
harnad@mind.UUCP (Stevan Harnad) (10/16/86)
In reply to the following by me in <167@mind.UUCP>: > there is no evidence at all that > either capacities or contexts are modular. michaelm@bcsaic.UUCP (michael maxwell) writes: >> Maybe I'm reading this out of context (not having read your books or papers), >> but could you explain this statement? I know of lots of evidence for the >> modularity of various aspects of linguistic behavior. In fact, we have a >> parser + grammar of English here that captures a large portion of English >> syntax, but has absolutely no semantics (yet). I'm afraid this extract is indeed a bit out of context. The original context concerned what I've dubbed the "Total Turing Test," one in which ALL of our performance capacities -- robotic and linguistic -- are "captured." In the papers under discussion I described several arguments in favor of the Total Turing Test over any partial turing test, such as "toy" models that only simulate a small chunk of our cognitive performance capacity, or even the (subtotal) linguistic ("teleteype") version of the Total Turing Test. These arguments included: (3) The "Convergence Argument" that `toy' problems are arbitrary, that they have too many degrees of freedom, that the d.f. shrink as the capacities of the toy grow to life-size, and that the only version that reduces the underdetermination to the normal proportions of a scientific theory is the `Total' one. (5) The "Nonmodularity Argument" that no subtotal model constitutes a natural module (insofar as the turing test is concerned); the only natural autonomous modules are other organisms, with their complete robotic capacities (more of this below). (7) The "Robotic Functionalist Argument" that the entire symbolic functional level is no macromodule either, and needs to be grounded in robotic function. I happen to have views on the "autonomy of syntax" (which is of course the grand-daddy of the current modulo-mania), but they're not really pertinent to the total vs modular turing-test issue. Perhaps the only point about an autonomous parser that is relevant here is that it is in the nature of the informal, intuitive component of the turing test that lifeless fragments of mimicry (such as Searle's isolated `thirst' module) are not viable; they simply fail to convince us of anything. And rightly so, I should think; otherwise the turing test would be a pretty flimsy one. Let me add, though, that even "convincing" autonomous parsing performance (in the non-turing sense of convincing) seems to me to be rather weak evidence for the psychological reality of a syntactic module -- let alone that it has a mind. (On my theory, semantic performance has to be grounded in robotic performance and syntactic performance must in turn be grounded in semantic performance.) Stevan Harnad (princeton!mind!harnad)
greid@adobe.UUCP (10/17/86)
It seems to me that the idea of concocting a universal Turing test is sort of useless. Consider, for a moment, monsters. There have been countless monsters on TV and film that have had varying degrees of human-ness, and as we watch the plot progress, we are sort of administering the Turing test. Some of the better aliens, like in "Blade Runner", are very difficult to detect as being non-human. However, given enough time, we will eventually notice that they don't sleep, or that they drink motor oil, or that they don't bleed when they are cut (think of "Terminator" and surgery for a minute), and we start to think of alternative explanations for the aberrances we have noticed. If we are watching TV, we figure it is a monster. If we are walking down the street and we see somebody get their arm cut off and they don't bleed, we think *we* are crazy (or we suspect "special effects" and start looking for the movie camera), because there is no other plausible explanation. There are even human beings whom we question when one of our subconscious "tests" fails--like language barriers, brain damage, etc. If you think about it, there are lots of human beings who would not pass the Turing test. Let's forget about it. Glenn Reid Adobe Systems Adobe claims no knowledge of anything in this message.
harnad@mind.UUCP (Stevan Harnad) (10/18/86)
In response to some of the arguments in favor of the robotic over the symbolic version of the turing test in (the summaries of) my articles "Minds, Machines and Searle" and "Category Induction and Representation" franka@mmintl.UUCP (Frank Adams) replies: > [R]elating purely symbolic functions to external events is > essentially a solved problem. Digital audio recording, for > example, works quite well. Robotic operations generally fail, > when they do, not because of any problems with the digital > control of an analog process, but because the purely symbolic > portion of the process is inadequate. In other words, there is > every reason to expect that a computer program able to pass the > [linguistic version of the] Turing test could be extended to one > able to pass the robotic version...requiring additional development > effort which is tiny by comparison (though likely still measured > in man-years). This argument has become quite familiar to me from delivering the oral version of the papers under discussion. It is the "Triviality of Transduction [A/D conversion, D/A conversion, Effectors] Argument" (TT for short). Among my replies to TT the central one is the principled Antimodularity Argument: There are reasons to believe that the neat partitioning of function into autonomous symbolic and nonsymbolic modules may break down in the special case of mind modeling. These reasons include my "Groundedness" Argument: that unless cognitive symbols are grounded (psychophysically, bottom-up) in nonsymbolic processes they remain meaningless. (This amounts to saying that we must be intrinsically "dedicated" devices and that our A/D and our "decryption/encryptions" are nontrivial; in passing, this is also a reply to Searle's worries about "intrinsic" versus "derived" intentionality. It may also be the real reason why "the purely symbolic portion of the process is inadequate"!) This problem of grounding symbolic processes in nonsymbolic ones in the special case of cognition is also the motivation for the material on category representation. Apart from nonmodularity and groundedness, other reasons include: (1) Searle's argument itself, and the fact that only the transduction argument can block it; that's some prima facie ground for believing that the TT may be false in the special case of mind-modeling. (2) The triviality of ordinary (nonbiological) transduction and its capabilities, comparared to what organisms with senses (and minds) can do. (Compare the I/O capacities of "audio" devices with those of "auditory" ones; the nonmodular road to the capacity to pass the total turing test suggests that we are talking here about qualitative differences, not quantitative ones.) (3) Induction (both ontogenetic and phylogentetic) and inductive capacity play an intrinsic and nontrivial role in bio-transduction that they do not play in ordinary engineering peripherals, or the kinds of I/O problems these have been designed for. (4) Related to the Simulation/Implementation Argument: There are always more real-world contingencies than can be anticipated in a symbolic description or simulation. That's why category representations are approximate and the turing test is open-ended. For all these reasons, I believe that Object/Symbol conversion in cognition is a considerably more profound problem than ordinary A/D; orders of magnitude more profound, in fact, and hence that TT is false. > [E]ven in a purely formal environment, there turn out to be a > lot of real things to talk about. Primitive concepts of time > (before and after) are understandable. One can talk about nouns > and verbs, sentences and conversations, self and other. I don't > see any fundamental difference between the ability to deal with > symbols as real objects, and the ability to deal with other kinds > of real objects. I don't completely understand the assumptions being made here. (What is a "purely formal environment"? Does anyone you know live in one?) Filling in with some educated guesses here, I would say that again the Object/Symbol conversion problem in the special case of organisms' mental capacities is being vastly underestimated. Object-manipulation (including discrimination, categorization, identification and description) is not a mere special case of symbol-manipulation or vice-versa. One must be grounded in the other in a principled way, and the principles are not yet known. On another interpretation, perhaps you are talking about "deixis" -- the necessity, even in the linguistic (symbolic) version of the turing test, to be able to refer to real objects in the here-and-now. I agree that this is a deep problem, and conjecture that its solution in the symbolic version will have to draw on anterior nonsymbolic (i.e., robotic) capacities. Stevan Harnad princeton!mind!harnad
harnad@mind.UUCP (Stevan Harnad) (10/19/86)
greid@adobe.UUCP (Glenn Reid) writes: > [C]oncocting a universal Turing test is sort of useless... There > have been countless monsters on TV...[with] varying degrees of > human-ness...Some...very difficult to detect as being non-human. > However, given enough time, we will eventually notice that they > don't sleep, or that they drink motor oil... The objective of the turing test is to judge whether the candidate has a mind, not whether it is human or drinks motor oil. We must accordingly consult our intuitions as to what differences are and are not relevant to such a judgment. [Higher animals, for example, have no trouble at all passing (the animal version) of the turing test as far as I'm concerned. Why should aliens, monsters or robots, if they have what it takes in the relevant respects? As I have argued before, turing-testing for relevant likeness is really our only way of contending with the "other-minds" problem.] > [T]here are lots of human beings who would not pass the Turing > test [because of brain damage, etc.]. And some of them may not have minds. But we give them the benefit of the doubt for humanitarian reasons anyway. Stevan Harnad (princeton!mind!harnad)
rggoebel@watdragon.UUCP (Randy Goebel LPAIG) (10/19/86)
Stevan Harnad writes: > ...The objective of the turing test is to judge whether the candidate > has a mind, not whether it is human or drinks motor oil. This stuff is getting silly. I doubt that it is possible to test whether something has a mind, unless you provide a definition of what you believe a mind is. Turing's test wasn't a test for whether or not some artificial or natural entity had a mind. It was his prescription for an evaluation of intelligence.
harnad@mind.UUCP (Stevan Harnad) (10/20/86)
rggoebel@watdragon.UUCP (Randy Goebel LPAIG) replies: > I doubt that it is possible to test whether something has a mind, > unless you provide a definition of what you believe a mind is. > Turing's test wasn't a test for whether or not some artificial > or natural entity had a mind. It was his prescription for an > evaluation of intelligence. And what do you think "having intelligence" is? Turing's criterion effectively made it: having performance capacity that is indistinguishable from human performance capacity. And that's all "having a mind" amounts to (by this objective criterion). There's no "definition" in any of this, by the way. We'll have definitions AFTER we have the functional answers about what sorts of devices can and cannot do what sorts of things, and how and why. For the time being all you have is a positive phenomenon -- having a mind, having intelligence -- and an objective and intuitive criterion for inferring its presence in any other case than one's own. (In your own case you presumable know what it's like to have-a-mind/have-intelligence on subjective grounds.) Stevan Harnad princeton!mind!harnad
michaelm@bcsaic.UUCP (10/20/86)
In article <1862@adobe.UUCP> greid@adobe.UUCP (Glenn Reid) writes: >...If you think >about it, there are lots of human beings who would not pass the Turing test. He must mean me, on a Monday morning :-) -- Mike Maxwell Boeing Advanced Technology Center ...uw-beaver!uw-june!bcsaic!michaelm
michaelm@bcsaic.UUCP (10/21/86)
>Stevan Harnad writes: > ...The objective of the turing test is to judge whether the candidate > has a mind, not whether it is human or drinks motor oil. In a related vein, if I recall my history correctly, the Turing test has been applied several times in history. One occasion was the encounter between the New World and the Old. I believe there was considerable speculation on the part of certain European groups (fueled, one imagines, by economic motives) as to whether the American Indians had souls. The (Catholic) church ruled that they did, effectively putting an end to the controversy. The question of whether they had souls was the historical equivalent to the question of whether they had mind and/or intelligence, I suppose. I believe the Turing test was also applied to oranguatans, although I don't recall the details (except that the orangutans flunked). As an interesting thought experiment, suppose a Turing test were done with a robot made to look like a human, and a human being who didn't speak English-- both over a CCTV, say, so you couldn't touch them to see which one was soft, etc. What would the robot have to do in order to pass itself off as human? -- Mike Maxwell Boeing Advanced Technology Center ...uw-beaver!uw-june!bcsaic!michaelm
harnad@mind.UUCP (Stevan Harnad) (10/23/86)
michaelm@bcsaic.UUCP (michael maxwell) writes: > I believe the Turing test was also applied to orangutans, although > I don't recall the details (except that the orangutans flunked)... > As an interesting thought experiment, suppose a Turing test were done > with a robot made to look like a human, and a human being who didn't > speak English-- both over a CCTV, say, so you couldn't touch them to > see which one was soft, etc. What would the robot have to do in order > to pass itself off as human? They should all three in principle have a chance of passing. For the orang, we would need to administer the ecologically valid version of the test. (I think we have reasonably reliable cross-species intuitions about mental states, although they're obviously not as sensitive as our intraspecific ones, and they tend to be anthropocentric and anthropomorphic -- perhaps necessarily so; experienced naturalists are better at this, just as cross-cultural ethnographic judgments depend on exposure and experience.) We certainly have no problem in principle with foreign speakers (the remarkable linguist, polyglot and bible-translator Kenneth Pike has a "magic show" in which, after less than an hour of "turing" interactions with a speaker of any of the [shrinking] number of languages he doesn't yet know, they are babbling mutually intelligibly before your very eyes), although most of us may have some problems in practice with such a feat, at least, without practice. Severe aphasics and mental retardates may be tougher cases, but there perhaps the orang version would stand us in good stead (and I don't mean that disrespectfully; I have an extremely high regard for the mental states of our fellow creatures, whether human or nonhuman). As to the robot: Well that's the issue here, isn't it? Can it or can it not pass the appropriate total test that its appropriate non-robot counterpart (be it human or ape) can pass? If so, it has a mind, by this criterion (the Total Turing Test). I certainly wouldn't dream of flunking either a human or a robot just because he/it didn't feel soft, if his/its total performance was otherwise turing indistinguishable. Stevan Harnad princeton!mind!harnad harnad%mind@princeton.csnet
freeman@spar.SPAR.SLB.COM (Jay Freeman) (10/24/86)
Possibly a more interesting test would be to give the computer direct control of the video bit map and let it synthesize an image of a human being.
harnad@mind.UUCP (Stevan Harnad) (10/26/86)
freeman@spar.UUCP (Jay Freeman) replies: > Possibly a more interesting test [than the robotic version of > the Total Turing Test] would be to give the computer > direct control of the video bit map and let it synthesize an > image of a human being. Manipulating digital "images" is still only symbol-manipulation. It is (1) the causal connection of the transducers with the objects of the outside world, including (2) any physical "resemblance" the energy pattern on the transducers may have to the objects from which they originate, that distinguishes robotic functionalism from symbolic functionalism and that suggests a solution to the problem of grounding the otherwise ungrounded symbols (i.e., the problem of "intrinsic vs. derived intentionality"), as argued in the papers under discussion. A third reason why internally manipulated bit-maps are not a new way out of the problems with the symbolic version of the turing test is that (3) a model that tries to explain the functional basis of our total performance capacity already has its hands full with anticipating and generating all of our response capacities in the face of any potential input contingency (i.e., passing the Total Turing Test) without having to anticipate and generate all the input contingencies themselves. In other words, its enough of a problem to model the mind and how it interacts successfully with the world without having to model the world too. Stevan Harnad {seismo, packard, allegra} !princeton!mind!harnad harnad%mind@princeton.csnet (609)-921-7771
freeman@spar.UUCP (10/27/86)
<*munch*> In article <12@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: > >freeman@spar.UUCP (Jay Freeman) replies: > >> Possibly a more interesting test [than the robotic version of >> the Total Turing Test] would be to give the computer >> direct control of the video bit map and let it synthesize an >> image of a human being. > > Manipulating digital "images" is still only symbol-manipulation. [...] Very well, let's equip the robot with an active RF emitter so it can jam the camera's electronics and impose whatever bit map it wishes, whether the camera likes it or not. Too silly? Very well, let's design a robot in the shape of a back projector, and let it create internally whatever representation of a human being it wishes the camera to see, and project it on its screen for the camera to pick up. Such a robot might do a tolerable job of interacting with other parts of the "objective" world, using robot arms and whatnot of more conventional design, so long as it kept them out of the way of the camera. Still too silly? Very well, let's create a vaguely anthropomorphic robot and equip its external surfaces with a complete covering of smaller video displays, so that it can achieve the minor details of human appearance by projection rather than by mechanical motion. (We can use a crude electronic jammer to limit the amount of detail that the camera can see, if necessary.) Well, maybe our model shop is good enough to do most of the details of the robot convincingly, so we'll only have to project subtle details of facial expression. Maybe just the eyes. Slightly more seriously, if you are going to admit the presence of electronic or mechanical devices between the subject under test and the human to be fooled, you must accept the possibility that the test subject will be smart enough to detect their presence and exploit their weaknesses. Returning to a more facetious tone, consider a robot that looks no more anthropomorphic than your vacuum cleaner, but that is possessed of moderate manipulative abilities and a good visual perceptive apparatus, and furthermore, has a Swiss Army knife. Before the test commences, the robot sneakily rolls up to the camera and removes the cover. It locates the connections for the external video output, and splices in a substitute connection to an external video source which it generates. Then it replaces the camera cover, so that everything looks normal. And a test time, the robot provides whatever image it wants the testers to see. A dumb robot might have no choice but to look like a human being in order to pass the test. Why should a smart one be so constrained? -- Jay Freeman
michaelm@bcsaic.UUCP (michael maxwell) (10/28/86)
In article <10@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: > >michaelm@bcsaic.UUCP (me) wrote: > >> As an interesting thought experiment, suppose a Turing test were done >> with a robot made to look like a human, and a human being who didn't >> speak English-- both over a CCTV, say, so you couldn't touch them to >> see which one was soft, etc. What would the robot have to do in order >> to pass itself off as human? > >...We certainly have no problem in principle with >foreign speakers (the remarkable linguist, polyglot and bible-translator >Kenneth Pike has a "magic show" in which, after less than an hour of "turing" >interactions with a speaker of any of the [shrinking] number of languages he >doesn't yet know, they are babbling mutually intelligibly before your very >eyes), although most of us may have some problems in practice with such a >feat, at least, without practice. Yes, you can do (I have done) such "magic shows" in which you begin to learn a language using just gestures + what you pick up of the language as you go along. It helps to have some training in linguistics, particularly field methods. The Summer Institute of Linguistics (of which Pike is President Emeritus) gives such classes. After one semester you too can give a magic show! I guess what I had in mind for the revised Turing test was not using language at all--maybe I should have eliminated the sound link (and writing). What in the way people behave (facial expressions, body language etc.) would cue us to the idea the one is a human and the other a robot? What if you showed pictures to the examinees--perhaps beautiful scenes, and revolting ones? This is more a test for emotions than for mind (Mr. Spock would probably fail). But I think that a lot of what we think of as human is tied up in this nonverbal/ emotional level. BTW, I doubt whether the number of languages Pike knows is shrinking because of these monolingual demonstrations (aka "magic shows") he's doing. After the tenth language, you tend to forget what the second or third language was-- much less what you learned! -- Mike Maxwell Boeing Advanced Technology Center ...uw-beaver!uw-june!bcsaic!michaelm
me@utai.UUCP (10/30/86)
In article <1@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: >In reply to a prior iteration D. Simon writes: > >> I fail to see what [your "Total Turing Test"] has to do with >> the Turing test as originally conceived, which involved measuring >> up AI systems against observers' impressions, rather than against >> objective standards... Moreover, you haven't said anything concrete >> about what this test might look like. > >How about this for a first approximation: We already know, roughly >speaking, what human beings are able to "do" -- their total cognitive >performance capacity: They can recognize, manipulate, sort, identify and >describe the objects in their environment and they can respond and reply >appropriately to descriptions. Get a robot to do that. When you think >he can do everything you know people can do formally, see whether >people can tell him apart from people informally. > "respond and reply appropriately to descriptions". Very nice. Should be a piece of cake to formalize--especially once you've formalized recognition, manipulation, identification, and description (and, let's face it, any dumb old computer can sort). This is precisely what I was wondering when I asked you what this total Turing test looks like. Apparently, you haven't the foggiest idea, except that it would test roughly the same things that the old-fashioned, informal, does-it-look-smart-or-doesn't-it Turing test checks. In fact, none of the criteria you have described above seems defineable in any sense other than by reference to standard Turing test results ("gee, it sure classified THAT element the way I would've!"). And if you WERE to define the entire spectrum of human behaviour in an objective fashion ("rule 1: answering, 'splunge!' to any question is hereby defined as an 'appropriate reply'"), how would you determine whether the objective definition is useful? Why, build a robot embodying it, and see if people consider it intelligent, of course! The illusion of a "total" Turing test, distinct from the old-fashioned, subjective variety, thus vanishes in a puff of empiricism. And forget the well-that's-the-way-Science-does-it argument. It won't wash --see below. >> I believe that people in general dodge the "other minds" problem >> simply by accepting as a convention that human beings are by >> definition intelligent. > >That's an artful dodge indeed. And do you think animals also accept such >conventions about one another? Philosophers, at least, seem to >have noticed that there's a bit of a problem there. Looking human >certainly gives us the prima facie benefit of the doubt in many cases, >but so far nature has spared us having to contend with any really >artful imposters. Wait till the robots begin giving our lax informal >turing-testing a run for its money. > I haven't a clue whether animals think, or whether you think, for that matter. This is precisely my point. I don't believe we humans have EVER solved the "other minds" problem, or have EVER used the Turing test, even to try to resolve the question of whether there exist "other minds". The fact that you would like us to have done so, thus giving you a justification for the use of the (informal part of) the Turing test (and the subsequent implicit basing of the formal part on the informal part--see above), doesn't make it so. This is where your scientific-empirical model for developing the "total" Turing test out of the original falls down. Let's examine the development of a typical scientific concept: You have some rough, intuitive observations of phenomena (gravity, stars, skin). You take some objects whose properties you believe you understand (rocks, telescopes, microscopes), let them interact with your vaguely observed phenomenon, and draw more rigorous conclusions based on the recorded results of these experimental interactions. Now, let's examine the Turing test in that light: we take possibly-intelligent robot R, whose properties are fairly well understood, and sit it in front of person P, whose properties are something of a cipher to us. We then have them interact, and get a reading off person P (such as, "yup, shore is smart", or, "nope, dumb as a tree"). Now, what properties are being scientifically investigated here? They can't have anything to do with robot R--we assume that R's designer, Dr. Rstein, already has a fairly good idea what R is about. Rather, it appears as though you are discerning those attributes of people which relate to their judgment of intelligence in other objects. Of course, it might well turn out that something productive comes out of this, but it's also quite possible (and I conjecture that it's actually quite likely) that what you get out of this is some scientific law such as, "anything which is physically indistinguishable from a human being and can mutter something that sounds like person P's language is intelligent; anything else is generally dumb, but possibly intelligent, depending on the decoration of the room and the drug content of P's bloodstream at the time of the test". In short, my worries about the context-dependence and subjective quality of the results have not disappeared in a puff of empiricism; they loom as large as ever. > >> WHAT DOES THE "TOTAL TURING TEST" LOOK LIKE?... Please >> forgive my impertinent questions, but I haven't read your >> articles, and I'm not exactly clear about what this "total" >> Turing test entails. > >Try reading the articles. > Well, not only did I consider this pretty snide, but when I sent you mail privately, asking politely where I can find the articles in question, I didn't even get an answer, snide or otherwise. So starting with this posting, I refuse to apologize for being impertinent. Nyah, nyah, nyah. > > >Stevan Harnad >princeton!mind!harnad Daniel R. Simon "sorry, no more quotations" -D. Simon
harnad@mind.UUCP (Stevan Harnad) (11/01/86)
Jay Freeman (freeman@spar.UUCP) had, I thought, joined the ongoing discussion about the robotic version of the Total Turing Test to address the questions that were raised in the papers under discussion, namely: (1) Do we have any basis for contending with the "other minds problem" -- whether in other people, animals or machines -- other than turing-indistinguishable performance capacity? (2) Is the teletype version of the turing test -- which allows only linguistic (i.e., symbolic) interactions -- a strong enough test? (3) Could even the linguistic version alone be successfully passed by any device whose symbolic functions were not "grounded" in nonsymbolic (i.e., robotic) function? (4) Are transduction, analog representations, A/D conversion, and effectors really trivial in this context, or is there a nontrivial hybrid function, grounding symbolic representation in nonsymbolic representation, that no one has yet worked out? When Freeman made his original sugestion that the symbolic processor could have access to the robotic transducer's bit-map, I thought he was making the sophisticated (but familiar) point that once the transducer representation is digitized, it's symbolic all the way. (This is a variant of the "transduction-is-trivial" argument.) My prior reply to Freeman (about simulated models of the world, modularity, etc.) was addressed to this construal of his point. But now I see that he was not making this point at all, for he replies: > ... let's equip the robot with an active RF emitter so > it can jam the camera's electronics and impose whatever bit map it > wishes... design a robot in the shape of a back projector, and let it > create internally whatever representation of a human being it wishes > the camera to see, and project it on its screen for the camera to > pick up. Such a robot might do a tolerable job of interacting with > other parts of the "objective" world, using robot arms and whatnot > of more conventional design, so long as it kept them out of the > way of the camera... let's create a vaguely anthropomorphic robot and > equip its external surfaces with a complete covering of smaller video > displays, so that it can achieve the minor details of human appearance > by projection rather than by mechanical motion. Well, maybe our model > shop is good enough to do most of the details of the robot convincingly, > so we'll only have to project subtle details of facial expression. > Maybe just the eyes. > ... if you are going to admit the presence of electronic or mechanical > devices between the subject under test and the human to be fooled, > you must accept the possibility that the test subject will be smart > enough to detect their presence and exploit their weaknesses... > consider a robot that looks no more anthropomorphic than your vacuum > cleaner, but that is possessed of moderate manipulative abilities and > a good visual perceptive apparatus. > Before the test commences, the robot sneakily rolls up to the > camera and removes the cover. It locates the connections for the > external video output, and splices in a substitute connection to > an external video source which it generates. Then it replaces the > camera cover, so that everything looks normal. And at test time, > the robot provides whatever image it wants the testers to see. > A dumb robot might have no choice but to look like a human being > in order to pass the test. Why should a smart one be so constrained? From this reply I infer that Freeman is largely concerned with the question of appearance: Can a robot that doesn't really look like a person SIMULATE looking like a person by essentially symbolic means, plus add-on modular peripherals? In the papers under discussion (and in some other iterations of this discussion on the net) I explicitly rejected appearance as a criterion. (The reasons are given elsewhere.) What is important in the robotic version is that it should be a human DO-alike, not a human LOOK-alike. I am claiming that the (Total) object-manipulative (etc.) performance of humans cannot be generated by a basically symbolic module that is merely connected with peripheral modules. I am hypothesizing (a) that symbolic representations must be NONMODULARLY (i.e., not independently) grounded in nonsymbolic representations, (b) that the Total Turing Test requires the candidate to display all of our robotic capacities as well as our linguistic ones, and (c) that even the linguistic ones could not be accomplished unless grounded in the robotic ones. In none of this do the particulars of what the robot (or its grey matter!) LOOK like matter. Two last observations. First, what the "proximal stimulus" -- i.e., the physical energy pattern on the transducer surface -- PRESERVES whereas the next (A/D) step -- the digital representation -- LOSES, is everything about the full PHYSICAL configuration of the energy pattern that cannot be recovered by inversion (D/A). (That's what the ongoing concurrent discussion about the A/D distinction is in part concerned with.) Second, I think there is a tendency to overcomplicate the issues involved in the turing test by adding various arbitrary elaborations to it. The basic questions are fairly simply stated (though not so simple to answer). Focusing instead on ornamented variants often seems to lead to begging the question or changing the subject. Stevan Harnad {allegra, bellcore, seismo, rutgers, packard} !princeton!mind!harnad harnad%mind@princeton.csnet (609)-921-7771
harnad@mind.UUCP (Stevan Harnad) (11/01/86)
michaelm@bcsaic.UUCP (michael maxwell) writes: > I guess what I had in mind for the revised Turing test was not using > language at all--maybe I should have eliminated the sound link (and > writing). What in the way people behave (facial expressions, body > language etc.) would cue us to the idea the one is a human and the other > a robot? What if you showed pictures to the examinees--perhaps > beautiful scenes, and revolting ones? This is more a test for emotions > than for mind (Mr. Spock would probably fail). But I think that a lot of > what we think of as human is tied up in this nonverbal/ emotional level. The modularity issue looms large again. I don't believe there's an independent module for affective expression in human beings. It's all -- to use a trendy though inadequate expression -- "cognitively penetrable." There's also the issue of the TOTALITY of the Total Turing Test, which was intended to remedy the underdetermination of toy models/modules: It's not enough just to get a model to mimic our facial expressions. That could all be LITERALLY done with mirrors (and, say, some delayed feedback and some scrambling and recombining), and I'm sure it could fool people, at least for a while. I simply conjecture that this could not be done for the TOTALITY of our performance capacity using only more of the same kinds of tricks (analog OR symbolic). The capacity to manipulate objects in the world in all the ways we can and do do it (which happens to include naming and describing them, i.e., linguistic acts) is a lot taller order than mimicking exclusively our nonverbal expressive behavior. There may be (in an unfortunate mixed metaphor) many more ways to skin (toy) parts of the theoretical cat than all of it. Three final points: (1) Your proposal seems to equivocate between the (more important) formal functional component of the Total Turing Test (i.e., how do we get a model to exhibit all of our performance capacities, be they verbal or nonverbal?) and (2) the informal, intuitive component (i.e., will it be indistinguishable in all relevant respects from a person, TO a person?). The motto would be: If you use something short of the Total Turing Test, you may be able to fool some people some of the time, but not all of the time. (2) There's nothing wrong in principle with a nonverbal, even a nonhuman turing test; I think (higher) animals pass this easily all the time, with virtually the same validity as humans, as far as I'm concerned. But this version can't rely exclusively on affective expression modules either. (3) Finally, as I've argued earlier, all attempts to "capture" qualitative experience -- not just emotion, but any conscious experience, such as what it's LIKE to see red or to believe X -- amounts to an unprofitable red herring in this enterprise. The whole point of the Total Turing Test is that performance-indistinguishability IS your only basis for infer that anyone but you has a mind (i.e., has emotions, etc.). In the paper I dubbed this "methodological epiphenomenalism as aresearch strategy in cognitive science." By the way, you prejudged the question the way you put it. A perfectly noncommittal but monistic way of putting it would be: "What in the way ROBOTS behave would cue us to the idea that one robot had a mind and another did not?" This leaves it appropriately open for continuing research just exactly which causal physical devices (= "robots"), whether natural or artificial, do or do not have minds. Stevan Harnad {allegra, bellcore, seismo, rutgers, packard} !princeton!mind!harnad harnad%mind@princeton.csnet (609)-921-7771
harnad@mind.UUCP (Stevan Harnad) (11/01/86)
In his second net.ai comment on the abstracts of the two articles under discussion, me@utai.UUCP (Daniel Simon) wrote: >> WHAT DOES THE "TOTAL TURING TEST" LOOK LIKE?... Please >> forgive my impertinent questions, but I haven't read your >> articles, and I'm not exactly clear about what this "total" >> Turing test entails. I replied (after longish attempts to explain in two separate iterations): >"Try reading the articles." Daniel Simon rejoined: > Well, not only did I consider this pretty snide, but when I sent you > mail privately, asking politely where I can find the articles in > question, I didn't even get an answer, snide or otherwise. So starting > with this posting, I refuse to apologize for being impertinent. > Nyah, nyah, nyah. The same day, the following email came from Daniel Simon: > Subject: Hoo, boy, did I put my foot in it: > Ooops....Thank you very much for sending me the articles, and I'm sorry > I called you snide in my last posting. If you see a bright scarlet glow > in the distance, looking west from Princeton, it's my face. Serves me > right for being impertinent in the first place... As soon as I finish > reading the papers, I'll respond in full--assuming you still care what > I have to say... Thanks again. Yours shamefacedly, Daniel R. Simon. This is a very new form of communication for all of us. We're just going to have to work out a new code of Nettiquette. With time, it'll come. I continue to care what anyone says with courtesy and restraint, and intend to respond to everything of which I succeed in making sense. Stevan Harnad {allegra, bellcore, seismo, rutgers, packard} !princeton!mind!harnad harnad%mind@princeton.csnet (609)-921-7771
harnad@mind.UUCP (Stevan Harnad) (11/02/86)
Here are backups of 5 prior replies that never made it to mod.ai. They are responses to Cugini, Kalish, Krulwich, Mozes and Paul. If you've read them elsewhere, please skip this file... Stevan Harnad ----- (1) In Message-ID: <8610190504.AA08059@ucbvax.Berkeley.EDU> on mod.ai CUGINI, JOHN <cugini@nbs-vms.ARPA> replies to my claim that >> there is no rational reason for being more sceptical about robots' >> minds (if we can't tell their performance apart from that of people) >> than about (other) peoples' minds. with the following: > One (rationally) believes other people are conscious BOTH because > of their performance and because their internal stuff is a lot like > one's own. This is a very important point and a subtle one, so I want to make sure that my position is explicit and clear: I am not denying that there exist some objective data that correlate with having a mind (consciousness) over and above performance data. In particular, there's (1) the way we look and (2) the fact that we have brains. What I am denying is that this is relevant to our intuitions about who has a mind and why. I claim that our intuitive sense of who has a mind is COMPLETELY based on performance, and our reason can do no better. These other correlates are only inessential afterthoughts, and it's irrational to take them as criteria. My supporting argument is very simple: We have absolutely no intuitive FUNCTIONAL ideas about how our brains work. (If we did, we'd have long since spun an implementable brain theory from our introspective armchairs.) Consequently, our belief that brains are evidence of minds and that the absence of a brain is evidence of the absence of a mind is based on a superficial black-box correlation. It is no more rational than being biased by any other aspect of appearance, such as the color of the skin, the shape of the eyes or even the presence or absence of a tail. To put it in the starkest terms possible: We wouldn't know what device was and was not relevantly brain-like if it was staring us in the face -- EXCEPT IF IT HAD OUR PERFORMANCE CAPACITIES (i.e., it could pass the Total Turing Test). That's the only thing our intuitions have to go on, and our reason has nothing more to offer either. To take one last pass at setting the relevant intuitions: We know what it's like to DO (and be able to do) certain things. Similar performance capacity is our basis for inferring that what it's like for me is what it's like for you (or it). We do not know anything about HOW we do any of those things, or about what would count as the right way and the wrong way (functionally speaking). Inferring that another entity has a mind is an intuitive judgment based on performance. It's called the (total) turing test. Inferring HOW other entities accomplish their performance is ordinary scientific inference. We're in no rational position to prejudge this profound and substantive issue on the basis of the appearance of a lump of grey jelly to our untutored but superstitious minds. > [W]e DO have some idea about the functional basis for mind, namely > that it depends on the brain (at least more than on the pancreas, say). > This is not to contend that there might not be other bases, but for > now ALL the minds we know of are brain-based, and it's just not > dazzlingly clear whether this is an incidental fact or somewhat > more deeply entrenched. The question isn't whether the fact is incidental, but what its relevant functional basis is. In other words, what is it about he brain that's relevant and what incidental? We need the causal basis for the correlation, and that calls for a hefty piece of creative scientific inference (probably in theoretical bio-engineering). The pancreas is no problem, because it can't generate the brain's performance capacities. But it is simply begging the question to say that brain-likeness is an EXTRA relevant source of information in turing-testing robots, when we have no idea what's relevantly brain-like. People were sure (as sure as they'll ever be) that other people had minds long before they ever discovered they had brains. I myself believed the brain was just a figure of speech for the first dozen or so years of my life. Perhaps there are people who don't learn or believe the news throughout their entire lifetimes. Do you think these people KNOW any less than we do about what does or doesn't have a mind? Besides, how many people do you think could really pick out a brain from a pancreas anyway? And even those who can have absolutely no idea what it is about the brain that makes it conscious; and whether a cow's brain or a horse-shoe crab's has it; or whether any other device, artificial or natural, has it or lacks it, or why. In the end everyone must revert to the fact that a brain is as a brain does. > Why is consciousness a red herring just because it adds a level > of uncertainty? Perhaps I should have said indeterminacy. If my arguments for performance-indiscernibility (the turing test) as our only objective basis for inferring mind are correct, then there is a level of underdetermination here that is in no way comparable to that of, say, the unobservable theoretical entities of physics (say, quarks, or, to be more trendy, perhaps strings). Ordinary underdetermination goes like this: How do I know that your theory's right about the existence and presence of strings? Because WITH them the theory succeeds in accounting for all the objective data (let's pretend), and without them it does not. Strings are not "forced" by the data, and other rival theories may be possible that work without them. But until these rivals are put forward, normal science says strings are "real" (modulo ordinary underdetermination). Now try to run that through for consciousness: How do I know that your theory's right about the existence and presence of consciousness (i.e., that your model has a mind)? "Because its performance is turing-indistinguishable from that of creatures that have minds." Is your theory dualistic? Does it give consciousness an independent, nonphysical, causal role? "Goodness, no!" Well then, wouldn't it fit the objective data just as well (indeed, turing-indistinguishably) without consciousness? "Well..." That's indeterminacy, or radical underdetermination, or what have you. And that's why consciousness is a methodological red herring. > Even though any correlations will ultimately be grounded on one side > by introspection reports, it does not follow that we will never know, > with reasonable assurance, which aspects of the brain are necessary for > consciousness and which are incidental...Now at some level of difficulty > and abstraction, you can always engineer anything with anything... But > the "multi-realizability" argument has force only if its obvious > (which it ain't) that the structure of the brain at a fairly high > level (eg neuron networks, rather than molecules), high enough to be > duplicated by electronics, is what's important for consciousness. We'll certainly learn more about the correlation between brain function and consciousness, and even about the causal (functional) basis of the correlation. But the correlation will really be between function and performance capacity, and the rest will remain the intuitive inference or leap of faith it always was. And since ascertaining what is relevant about brain function and what is incidental cannot depend simply on its BEING brain function, but must instead depend, as usual, on the performance criterion, we're back where we started. (What do you think is the basis for our confidence in introspective reports? And what are you going to say about robots'introspective reports...?) I don't know what you mean, by the way, about always being able to "engineer anything with anything at some level of abstraction." Can anyone engineer something to pass the robotic version of the Total Turing Test right now? And what's that "level of abstraction" stuff? Robots have to do their thing in the real world. And if my groundedness arguments are valid, that ain't all done with symbols (plus add-on peripheral modules). Stevan Harnad ----- (2) In mod.ai, Message-ID: <861016-071607-4573@Xerox>, "charles_kalish.EdServices"@XEROX.COM writes: > About Stevan Harnad's two kinds of Turing tests [linguistic > vs. robotic]: I can't really see what difference the I/O methods > of your system makes. It seems that the relevant issue is what > kind of representation of the world it has. I agree that what's at issue is what kind of representation of the world the system has. But you are prejudging "representation" to mean only symbolic representation, whereas the burden of the papers in question is to show that symbolic representations are "ungrounded" and must be grounded in nonsymbolic processes (nonmodularly -- i.e., NOT by merely tacking on autonomous peripherals). > While I agree that, to really understand, the system would need some > non-purely conventional representation (not semantic if "semantic" > means "not operable on in a formal way" as I believe [given the brain > is a physical system] all mental processes are formal then "semantic" > just means governed by a process we don't understand yet), giving and > getting through certain kinds of I/O doesn't make much difference. "Non-purely conventional representation"? Sounds mysterious. I've tried to make a concrete proposal as to just what that hybrid representation should be like. "All mental processes are formal"? Sounds like prejudging the issue again. It may help to be explicit about what one means by formal/symbolic: Symbolic processing is the manipulation of (arbitrary) physical tokens in virtue of their shape on the basis of formal rules. This is also called syntactic processing. The formal goings-on are also "semantically interpretable" -- they have meanings; they are connected to objects in the outside world that they are about. The Searle problem is that so far the only devices that do semantic interpretations intrinsically are ourselves. My proposal is that grounding the representations nonmodularly in the I/O connection may provide the requisite intrinsic semantics. This may be the "process we don't understand yet." But it means giving up the idea that "all mental processes are formal" (which in any case does not follow, at least on the present definition of "formal," from the fact that "the brain is a physical system"). > Two for instances: SHRDLU operated on a simulated blocks world. The > modifications to make it operate on real blocks would have been > peripheral and not have affected the understanding of the system. This is a variant of the "Triviality of Transduction (& A/D, & D/A, and Effectors)" Argument (TT) that I've responded to in another iteration. In brief, it's toy problems like SHRDLU that are trivial. The complete translatability of internal symbolic descriptions into the objects they stand for (and the consequent partitioning of the substantive symbolic module and the trivial nonsymbolic peripherals) may simply break down, as I predict, for life-size problems approaching the power to pass the Total Turing Test. To put it another way: There is a conjecture implicit in the solutions to current toy/microworld problems, namely, that something along essentially the same lines will suitably generalize to the grown-up/macroworld problem. What I'm saying amounts to a denial of that conjecture, with reasons. It is not a reply to me to simply restate the conjecture. > Also, all systems take analog input and give analog output. Most receive > finger pressure on keys and return directed streams of ink or electrons. > It may be that a robot would need more "immediate" (as opposed to > conventional) representations, but it's neither necessary nor sufficient > to be a robot to have those representations. The problem isn't marrying symbolic systems to any old I/O. I claim that minds are "dedicated" systems of a particular kind: The kind capable of passing the Total Turing Test. That's the only necessity and sufficiency in question. And again, the mysterious word "immediate" doesn't help. I've tried to make a specific proposal, and I've accepted the consequences, namely, that it's just not going to be a "conventional" marriage at all, between a (substantive) symbolic module and a (trivial) nonsymbolic module, but rather a case of miscegenation (or a sex-change operation, or some other suitably mixed metaphor). The resulting representational system will be grounded "bottom-up" in nonsymbolic function (and will, I hope, display the characteristic "hybrid vigor" that our current pure-bred symbolic and nonsymbolic processes lack), as I've proposed (nonmetaphorically) in the papers under discussion. Stevan Harnad ----- (3) KRULWICH@C.CS.CMU.EDU (Bruce Krulwich) writes: > i disagree...that symbols, and in general any entity that a computer > will process, can only be dealt with in terms of syntax. for example, > when i add two integers, the bits that the integers are encoded in are > interpreted semantically to combine to form an integer. the same > could be said about a symbol that i pass to a routine in an > object-oriented system such as CLU, where what is done with > the symbol depends on it's type (which i claim is it's semantics) Syntax is ordinarily defined as formal rules for manipulating physical symbol tokens in virtue of their (arbitrary) SHAPES. The syntactic goings-on are semantically interpretable, that is, the symbols are also manipulable in virtue of their MEANINGS, not just their shapes. Meaning is a complex and ill-understood phenomenon, but it includes (1) the relation of the symbols to the real objects they "stand for" and (2) a subjective sense of understanding that relation (i.e., what Searle has for English and lacks for Chinese, despite correctly manipulating its symbols). So far the only ones who seem to do (1) and (2) are ourselves. Redefining semantics as manipulating symbols in virtue of their "type" doesn't seem to solve the problem... > i think that the reason that computers are so far behind the > human brain in semantic interpretation and in general "thinking" > is that the brain contains a hell of a lot more information > than most computer systems, and also the brain makes associations > much faster, so an object (ie, a thought) is associated with > its semantics almost instantly. I'd say you're pinning a lot of hopes on "more" and "faster." The problem just might be somewhat deeper than that... Stevan Harnad ----- (4) On mod.ai, in Message-ID: <8610160605.AA09268@ucbvax.Berkeley.EDU> on 16 Oct 86 06:05:38 GMT, eyal@wisdom.BITNET (Eyal mozes) writes: > I don't see your point at all about "categorical > perception". You say that "differences between reds and differences > between yellows look much smaller than equal-sized differences that > cross the red/yellow boundary". But if they look much smaller, this > means they're NOT "equal-sized"; the differences in wave-length may be > the same, but the differences in COLOR are much smaller. There seems to be a problem here, and I'm afraid it might be the mind/body problem. I'm not completely sure what you mean. If all you mean is that sometimes equal-sized differences in inputs can be made unequal by internal differences in how they are encoded, embodied or represented -- i.e., that internal physical differences of some sort may mediate the perceived inequalities -- then I of course agree. There are indeed innate color-detecting structures. Moreover, it is the hypothesis of the paper under discussion that such internal categorical representations can also arise as a consequence of learning. If what you mean, however, is that there exist qualitative differences among equal-sized input differences with no internal physical counterpart, and that these are in fact mediated by the intrinsic nature of phenomenological COLOR -- that discontinuous qualitative inequalities can occur when everything physical involved, external and internal, is continuous and equal -- then I am afraid I cannot follow you. My own position on color quality -- i.e., "what it's like" to experience red, etc. -- is that it is best ignored, methodologically. Psychophysical modeling is better off restricting itself to what we CAN hope to handle, namely, relative and absolute judgments: What differences can we tell apart in pairwise comparison (relative discrimination) and what stimuli or objects can we label or identify (absolute discrimination)? We have our hands full modeling this. Further concerns about trying to capture the qualitative nature of perception, over and above its performance consequences [the Total Turing Test] are, I believe, futile. This position can be dubbed "methodological epiphenomenalism." It amounts to saying that the best empirical theory of mind that we can hope to come up with will always be JUST AS TRUE of devices that actual have qualitative experiences (i.e., are conscious) as of devices that behave EXACTLY AS IF they had qualitative experiences (i.e., turing-indistinguishably), but do not (if such insentient look-alikes are possible). The position is argued in detail in the papers under discussion. > Your whole theory is based on the assumption that perceptual qualities > are something physical in the outside world (e.g., that colors ARE > wave-lengths). But this is wrong. Perceptual qualities represent the > form in which we perceive external objects, and they're determined both > by external physical conditions and by the physical structure of our > sensory apparatus; thus, colors are determined both by wave-lengths and > by the physical structure of our visual system. So there's no apriori > reason to expect that equal-sized differences in wave-length will lead > to equal-sized differences in color, or to assume that deviations from > this rule must be caused by internal representations of categories. And > this seems to completely cut the grounds from under your theory. Again, there is nothing for me to disagree with if you're saying that perceived discontinuities are mediated by either external or internal physical discontinuities. In modeling the induction and representation of categories, I am modeling the physical sources of such discontinuities. But there's still an ambiguity in what you seem to be saying, and I don't think I'm mistaken if I think I detect a note of dualism in it. It all hinges on what you mean by "outside world." If you only mean what's physically outside the device in question, then of course perceptual qualities cannot be equated with that. It's internal physical differences that matter. But that doesn't seem to be all you mean by "outside world." You seem to mean that the whole of the physical world is somehow "outside" conscious perception. What else can you mean by the statement that "perceptual qualities represent the form [?] in which we perceive external objects" or that "there's no...reason to expect that...[perceptual] deviations from [physical equality]...must be caused by internal representations of categories." Perhaps I have misunderstood, but either this is just a reminder that there are internal physical differences one must take into account too in modeling the induction and representation of categories (but then they are indeed taken into account in the papers under discussion, and I can't imagine why you would think they would "completely cut the ground from under" my theory) or else you are saying something metaphysical with which I cannot agree. One last possibility may have to do with what you mean by "representation." I use the word eclectically, especially because the papers are arguing for a hybrid representation, with the symbolic component grounded in the nonsymbolic. So I can even agree with you that I doubt that mere symbolic differences are likely to be the sole cause of psychophysical discontinuities, although, being physically embodied, they are in principle sufficient. I hypothesize, though, that nonsymbolic differences are also involved in psychophysical discontinuities. > My second criticism is that, even if "categorical perception" really > provided a base for a theory of categorization, it would be very > limited; it would apply only to categories of perceptual qualities. I > can't see how you'd apply your approach to a category such as "table", > let alone "justice". How abstract categories can be grounded "bottom-up" in concrete psychophysical categories is the central theme of the papers under discussion. Your remarks were based only on the summaries and abstracts of those papers. By now I hope the preprints have reached you, as you requested, and that your question has been satisfactorily answered. To summarize "grounding" briefly: According to the model, (learned) concrete psychophysical categories are formed from sampling positive and negative instances of a category and then encoding the invariant information that will reliably identify further instances. This might be how one learned the concrete categories "horse" and "striped" for example. The (concrete) category "zebra" could then be learned without need for direct perceptual ACQUAINTANCE with the positive and negative instances by simply being told that a zebra is a striped horse. That is, the category can be learned by symbolic DESCRIPTION by merely recombining the labels of the already-grounded perceptual categories. All categorization involves some abstraction and generalization (even "horse," and certainly "striped" did), so abstract categories such as "goodness," "truth" and "justice" could be learned and represented by recursion on already grounded categories, their labels and their underlying representations. (I have no idea why you think I'd have a problem with "table.") > Actually, there already exists a theory of categorization that is along > similar lines to your approach, but integrated with a detailed theory > of perception and not subject to the two criticisms above; that is the > Objectivist theory of concepts. It was presented by Ayn Rand... and by > David Kelley... Thanks for the reference, but I'd be amazed to see an implementable, testable model of categorization performance issue from that source... Stevan Harnad ----- (5) Machines: Natural and Man-Made To Daniel Simon's reply in AI digest (V4 #226): >One question you haven't addressed is the relationship between intelligence and >"human performance". Are the two synonymous? If so, why bother to make >artificial humans when making natural ones is so much easier (not to mention >more fun)? Daniel Paul adds: > This is a question that has been bothering me for a while. When it > is so much cheaper (and possible now, while true machine intelligence > may be just a dream) why are we wasting time training machines when we > could be training humans instead? The only reasons that I can see are > that intelligent systems can be made small enough and light enough to > sit on bombs. Are there any other reasons? Apart from the two obvious ones -- (1) so machines can free people to do things machines cannot yet do, if people prefer, and (2) so machines can do things that people can only do less quickly and efficiently, if people prefer -- there is the less obvious reply already made to Daniel Simon: (3) because trying to get machines to display all our performance capacity (the Total Turing Test) is our only way of arriving at a functional understanding of what kinds of machines we are, and how we work. [Before the cards and letters pour in to inform me that I've used "machine" incoherently: A "machine," (writ large, Deus Ex Machina) is just a physical, causal system. Present-generation artificial machines are simply very primitive examples.] Stevan Harnad
harnad@mind.UUCP (Stevan Harnad) (11/03/86)
The following is a response on net.ai to a comment on mod.ai. Because of problems with posting to mod.ai, I am temporarily replying to net.ai. On mod.ai cugini@NBS-VMS.ARPA ("CUGINI, JOHN") writes: > You seem to want to pretend that we know absolutely nothing about the > basis of thought in humans, and to "suppress" all evidence based on > such knowledge. But that's just wrong. Brains *are* evidence for mind, > in light of our present knowledge. What I said was that we knew absolutely nothing about the FUNCTIONAL basis of thought in humans, i.e., about how brains or relevantly similar devices WORK. Hence we wouldn't have the vaguest idea if a given lump of grey matter was in fact the right stuff, or just a gelatenous look-alike -- except by examining its performance (i.e., turing) capacity. [The same is true, by the way, mutatis mutandis, for a better structural look-alike -- with cells, synapses, etc. We have no functional idea of what differentiates a mind-supporting look-alike from a comatose one, or one from a nonviable fetus. Without the performance criterion the brain cue could lead us astray as often as not regarding whether there was indeed a mind there. And that's not to mention that we knew perfectly well (perhaps better, even) how to judge whether somebody had a mind before 'ere we ope'd a skull nor knew what we had chanced upon there. If you want a trivial concession though, I'll make one: If you saw an inert body totally incapable of behavior, then or in the future, and you entertained some prior subjective probability that it had a mind, say, p, then, if you opened its skull and found something anatomically and physiologically brain-like in there, then the probability p that it had, or had had, a mind would correspondingly rise. Ditto for an inert alien species. And I agree that that would be rational. However, I don't think that any of that has much to do with the problem of modeling the mind, or with the relative strengths or weaknesses of the Total Turing Test. > People in, say, 1500 AD were perfectly rational in predicting > tides based on the position of the moon (and vice-versa) > even though they hadn't a clue as to the mechanism of interaction. > If you keep asking "why" long enough, *all* science is grounded on > such brute-fact correlation (why do like charges repel, etc.) - as > Hume pointed out a while back. Yes, but people then and even earlier were just as good at "predicting" the presence of mind WITHOUT any reference to the brain. And in ambiguous cases, behavior was and is the only rational arbiter. Consider, for example, which way you'd go if (1) an alien body persisted in behaving like a clock-like automaton in every respect -- no affect, no social interaction, just rote repetition -- but it DID have something that was indistinguishable (on the minute and superficial information we have) from a biological-like nervous system), versus (2) if a life-long close friend of yours had to undergo his first operation, and when they opened him up, he turned out to be all transistors on the inside. I don't set much store by this hypothetical sci-fi stuff, especially because it's not clear whether the "possibilities" we are contemplating are indeed possible. But the exercise does remind us that, after all, performance capacity is our primary criterion, both logically and intuitively, and its black-box correlates have whatever predictive power they may have only as a secondary, derivative matter. They depend for their validation on the behavioral criterion, and in cases of conflict, behavior continues to be the final arbiter. I agree that scientific inference is grounded in observed correlations. But the primary correlation in this special case is, I am arguing, between mental states and performance. That's what both our inferences and our intuitions are grounded in. The brain correlate is an additional cue, but only inasmuch as it agrees with performance. As to CAUSATION -- well, I'm sceptical that anyone will ever provide a completely satisfying account of the objective causes of subjective effects. Remember that, except for the special case of the mind, all other scientific inferences have only had to account for objective/objective correlations (and [or, more aptly, via) their subjective/subjective experiential counterparts). The case under discussion is the first (and I think only) case of objective/subjective correlation and causation. Hence all prior bets, generalizations or analogies are off or moot. > other brains... are, by definition, relevantly brain-like I'd be interested in knowing what current definition will distinguish a mind-supporting brain from a non-mind-supporting brain, or even a pseudobrain. (That IS the issue, after all, in claiming that the brain in an INDEPENDENT predictor of mindedness.) > Let me re-cast Harnad's argument (perhaps in a form unacceptable to > him): We can never know any mind directly, other than our own, if we > take the concept of mind to be something like "conscious intelligence" - > ie the intuitive (and correct, I believe) concept, rather than > some operational definition, which has been deliberately formulated > to circumvent the epistemological problems. (Harnad, to his credit, > does not stoop to such positivist ploys.) But the only external > evidence we are ever likely to get for "conscious intelligence" > is some kind of performance. Moreover, the physical basis for > such performance will be known only contingently, ie we do not > know, a priori, that it is brains, rather than automatic dishwashers, > which generate mind, but rather only as an a posteriori correlation. > Therefore, in the search for mind, we should rely on the primary > criterion (performance), rather than on such derivative criteria > as brains. I pretty much agree with the above account except for the > last sentence which prohibits us from making use of derivative > criteria. Why should we limit ourselves so? Since when is that part > of rationality? I accept the form in which you've recast my argument. The reasons that brainedness is not a good criterion are the following (I suppose I should stop saying it is not a "rational" criterion having made the minor concession I did above): Let's call being able to pass the Total Turing Test the "T" correlate of having a mind, and let's call having a brain the "B" correlate. (1) The validity of B depends completely on T. We have intuitions about the way we and others behave, and what it feels like; we have none about having brains. (2) In case of conflict between T and B, our intuitions (rationally, I suggest) go with T rather than B. (3) The subjective/objective issue (i.e., the mind/body problem) mentioned above puts these "correlations" in a rather different category from other empirical correlations, which are uniformly objective/objective. (4) Looked at sufficiently minutely and functionally, we don't know what the functionally relevant as opposed to the superficial properties of a brain are, insofar as mind-supportingness is concerned; in other words, we don't even know what's a B and what's just naively indistinguishable from a B (this is like a caricature of the turing test). Only T will allow us to pick them out. I think those are good enough reasons for saying that B is not a good independent criterion. That having be said, let me concede that for a radical sceptic, neither is T, for pretty pretty much the same reasons! This is why I am a methodological epiphenomenalist. > No, the fact is we do have more reason to suppose mind of other > humans than of robots, in virtue of an admittedly derivative (but > massively confirmed) criterion. And we are, in this regard, in an > epistemological position *superior* to those who don't/didn't know > about such things as the role of the brain, ie we have *more* reason > to believe in the mindedness of others than they do. That's why > primitive tribes (I guess) make the *mistake* of attributing > mind to trees, weather, etc. Since raw performance is all they > have to go on, seemingly meaningful activity on the part of any > old thing can be taken as evidence of consciousness. But we > sophisticates have indeed learned a thing or two, in particular, that > brains support consciousness, and therefore we (rationally) give the > benefit of the doubt to any brained entity, and the anti-benefit to > un-brained entities. Again, not to say that we might not learn about > other bases for mind - but that hardly disparages brainedness as a > rational criterion for mindedness. A trivially superior position, as I've suggested. Besides, the primitive's mistake (like the toy AI-modelers') is in settling for anything less than the Total Turing Test; the mistake is decidedly NOT the failure to hold out for the possession of a brain. I agree that it's rational to take brainedness as an additional corroborative cue, if you ever need one, but since it's completely useless when it fails to corroborate or conflicts with the Total Turing criterion, of what independent use is it? Perhaps I should repeat that I take the context for this discussion to be science rather than science fiction, exobiology or futurology. The problem we are presumably concerned with is that of providing an explanatory model of the mind along the lines of, say, physics's explanatory model of the universe. Where we will need "cues" and "correlates" is in determining whether the devices we build have succeeded in capturing the relevant functional properties of minds. Here the (ill-understood) properties of brains will, I suggest, be useless "correlates." (In fact, I conjecture that theoretical neuroscience will be led by, rather than itself leading, theoretical "mind-science" [= cognitive science?].) In sci-fi contexts, where we are guessing about aliens' minds or those of comatose creatures, having a blob of grey matter in the right place may indeed be predictive, but in the cog-sci lab it is not. > there's really not much difference between relying on one contingent > correlate (performance) rather than another (brains) as evidence for > the presence of mind. To a radical sceptic, as I've agreed above. But there is to a working cognitive scientist (whose best methodological stance, I suggest, is epiphenomenalism). > I know consciousness (my own, at least) exists, not as > some derived theoretical construct which explains low-level data > (like magnetism explains pointer readings), but as the absolutely > lowest rock-bottom datum there is. Consciousness is the data, > not the theory - it is the explicandum, not the explicans (hope > I got that right). It's true that I can't directly observe the > consciousness of others, but so what? That's an epistemological > inconvenience, but it doesn't make consciousness a red herring. I agree with most of this, and it's why I'm not, for example, an "eliminative materialist." But agreeing that consciousness is data rather than theory does not entail that it's the USUAL kind of data of empirical science. I KNOW I have a mind. Every other instance is radically different from this unique one: I can only guess, infer. Do you know of any similar case in normal scientific inference? This is not just an "epistemological inconvenience," it's a whole 'nother ball game. If we stick to the standard rules of objective science (which I recommend), then turing-indistinguishable performance modeling is indeed the best we can aspire to. And that does make consciousness a red herring. > ...being-composed-of-protein might not be as practically incidental > as many assume. Frinstance, at some level of difficulty, one can > get energy from sunlight "as plants do." But the issues are: > do we get energy from sunlight in the same way? How similar do > we demand that the processes are?...if we're interested in simulation at > a lower level of abstraction, eg, photosynthesis, then, maybe, a > non-biological approach will be impractical. The point is we know we > can simulate human chess-playing abilities with non-biological > technology. Should we just therefore declare the battle for mind won, > and go home? Or ask the harder question: what would it take to get a > machine to play a game of chess like a person does, ie, consciously. This sort of objection to a toy problem like chess (an objection I take to be valid) cannot be successfully redirected at the Total Turing Test, and that was one of the central points of the paper under discussion. Nor are the biological minutiae of modeling plant photosynthesis analogous to the biological minutiae of modeling the mind: The OBJECTIVE data in the mind case are what you can observe the organism to DO. Photosynthesis is something a plant does. In both cases one might reasonably demand that a veridical model should mimic the data as closely as possible. Hence the TOTAL Turing Test. But now what happens when you start bringing in physiological data, in the mind case, to be included with the performance data? There's no duality in the case of photosynthesis, nor is there any dichotomy of levels. Aspiring to model TOTAL photosynthesis is aspiring to get every chemical and temporal detail right. But what about the mind case? On the one hand, we both agree with the radical sceptic that NEITHER mimicking the behavior NOR mimicking the brain can furnish "direct" evidence that you've captured mind. So whereas getting every (observable) photosynthetic detail right "guarantees" that you've captured photosynthesis, there's no such guarantee with consciousness. So there's half of the disanalogy. Now consider again the hypothetical possibilities we were considering earlier: What if brain data and behavioral data compete? Which way should a nonsceptic vote? I'd go with behavior. Besides, it's an empirical question, as I said in the papers under discussion, whether or not brain constraints turn out to be relevant on the way to Total Turing Utopia. Way down the road, after all, the difference between mind-performance and brain-performance may well become blurred. Or it may not. I think the Total Turing Test is the right provisional methodology for getting you there, or at least getting you close enough. The rest may very well amount to only the "fine tuning." > BTW, I quite agree with your more general thesis on the likely > inadequacy of symbols (alone) to capture mind. I'm glad of that. But I have to point out that a lot of what you appear to disagree about went into the reasons supporting that very thesis, and vice versa. ----- May I append here a reply to andrews@ubc-cs.UUCP (Jamie Andrews) who wrote: > This endless discussion about the Turing Test makes the > "eliminative materialist" viewpoint very appealing: by the > time we have achieved something that most people today would > call intelligent, we will have done it through disposing of > concepts such as "intelligence", "consciousness", etc. > Perhaps the reason we're having so much trouble defining > a workable Turing Test is that we're essentially trying to > fit a square peg into a round hole, belabouring some point > which has less relevance than we realize. I wonder what old > Alan himself would say about the whole mess. On the contrary, rather than disposing of them, we will finally have some empirical and theoretical idea of what their functional basis might be, rather than simply knowing what it's like to have them. And if we don't first sort out our methodological constraints, we're not headed anywhere but in hermeneutic circles. Stevan Harnad {allegra, bellcore, seismo, rutgers, packard} !princeton!mind!harnad harnad%mind@princeton.csnet (609)-921-7771
kgd@rlvd.UUCP (Keith Dancey) (11/03/86)
In article <5@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes: > > >What do you think "having intelligence" is? Turing's criterion >effectively made it: having performance capacity that is indistinguishable >from human performance capacity. And that's all "having a mind" >amounts to (by this objective criterion). ... At the risk of sidetracking this discussion, I don't think it wise to try and equate 'mind' and 'intelligence'. A 'mind' is an absolute thing, but 'intelligence' is relative. For instance, most people would, I believe, accept that a monkey has a 'mind'. However, they would not necessarily so easily accept that a monkey has 'performance capacity that is indistinguishable from human performance capacity'. On the other hand, many people would accept that certain robotic processes had 'intelligence', but would be very reluctant to attribute them with 'minds'. I think there is something organic about 'minds', but 'intelligence' can be codified, within limits, of course. I apologise if this appears as a red-herring in the argument. -- Keith Dancey, UUCP: ..!mcvax!ukc!rlvd!kgd Rutherford Appleton Laboratory, Chilton, Didcot, Oxon OX11 0QX JANET: K.DANCEY@uk.ac.rl Tel: (0235) 21900 ext 5716