From Article: 5743 in comp.ai.philosophy Newsgroups: comp.ai,comp.ai.philosophy,comp.ai.nat-lang,sci.lang,sci.cognitive From: A.Sloman@cs.bham.ac.uk (Aaron Sloman) Subject: Re: Origin of "Symbol Grounding Problem"? (Long) Message-ID: Sender: news@cs.bham.ac.uk Nntp-Posting-Host: fattie Organization: School of Computer Science, University of Birmingham, UK References: <256mti$90h@molly.anu.edu.au> Date: Sat, 4 Sep 1993 18:32:01 GMT rapaport@cs.buffalo.edu (William J. Rapaport) writes: > Organization: State University of New York at Buffalo/Comp Sci > Date: Tue, 31 Aug 1993 16:51:52 GMT > > In article <256mti$90h@molly.anu.edu.au> hjc607@huxley.anu.edu.au (Hugh J Clapin) writes: > >Geoffrey Simmons (simmons@bosun1.informatik.uni-hamburg.de) wrote: > >: Does anybody know where the "symbol grounding problem" originated? That is, > >: where was it first stated as a problem under this name? (AI people are most > >: likely to have heard of it.) > > Harnad, Stevan (1990), ``The Symbol Grounding Problem,'' > _Physica D_ 42: 335--346 The problem, though not the name, is much older. For instance, Searle's attack on what he called the Strong AI thesis John Searle, `Minds Brains and Programs' in The Behavioural and Brain Sciences, 3,3 1980. I think that John Haugeland previously referred to it as the problem whether a machine could have "original" (non-derivative) intentionality, but I don't have a reference handy. In case anyone is interested I've discussed at some length what I think is essentially this problem, without using the phrase, in two papers in 1985 and 1986. `What enables a machine to understand?' in Proceedings 9th International Joint Conference on AI, pp 995-1001, Los Angeles, August 1985. `Reference without causal links' in Proceedings 7th European Conference on Artificial Intelligence, Brighton, July 1986. The proceedings were re-printed as J.B.H. du Boulay, D.Hogg, L.Steels (eds) Advances in Artificial Intelligence - II North Holland, 369-381, 1987. Summary: Many people (e.g. Harnad) hold the mistaken view that it's a single all-or-nothing matter whether an organism or machine can "really" use symbols to refer to things, or can interpret or understand symbols. This tends to be linked with the question whether the organism or machine can have experiences. E.g. Harnard assumes we can always ask of any machine or organism "Is there anybody home?" assuming there must always be a yes or no answer. This is a mistake. Our symbol understanding capability is a complex mixture of different sub-capabilities, not all of which need be present or absent together. Thus different organisms and machines may possess different subsets of these capabilities, with very different consequences. To understand these matters we need to survey and distinguish all these capabilities, and to investigate architectures and mechanisms that are capable of supporting different combinations of capabilities. (A particularly simple case, discussed in the 1985 paper, is the ability of a typical computer without any AI capabilities to use bit patterns (e.g. in an address register) to refer to locations in its address space. The argument that it's a human designer who is using the bit patterns is totally implausible given that no human may be aware of the particular combinations of bit patterns or the purpose for which they are actually being used at any time, e.g. in an automatic garbage collection program. A slightly more complex capability, which can be implemented on top of bit-manipulating capabilities, is the ability of a computer with AI software to build descriptions of some of the contents of its memory, and to check whether such a description is true or false, and in the latter case to synthesis a plan to make it true. These capabilities can exist without any ability to refer to anything *outside* the computer. That requires a richer architecture, as discussed in the 1986 paper. A still richer architecture is required for the machine to have its own desires and to use internal symbols in serving those desires. I am still working on those requirements!) If we pay attention to these issues, then, instead of a simple dichotomy between those things that CAN and those that CANNOT use symbols to refer we'll find something far more interesting: a wide range of cases combining different features. Instead of trying to categorise them using ordinary language or traditional philosophical concepts (e.g. "understand", "experience", "intentionality", "anybody home"), which are not rich enough for the task, we should try to develop a theory-based taxonomy of cases (much as the development of a theory of the mechanisms and structures in physical matter led to a theory-based table of types of elements: the periodic table.) Note that my claim (call it "A") that there's a range of different cases based on different subsets of capabilities is not the same as and does not support the claim (call it "B") that there's a *continuous spectrum* of cases, which is a tempting but mistaken response to people like Harnad. (B) implies that there are no discontinuities, so that boundaries are arbitrary (like the division between hills and mountains, perhaps?). (A) implies there are *many* discontinuities, corresponding to the presence or absence of particular sub-capabilities: e.g. -- the presence or absence of the ability to combine representations using truth-functional connectives; or -- the ability to use variable-binding constructs like universal and existential quantifiers; or -- the ability to change symbols continuously with continuously changing interpretations, as in diagrams; or -- the ability to use the same symbols in different roles, such as expressing beliefs, expressing desires, expressing suppositions, expressing parts of plans, and so on. Analysing these different capabilities, their requirements, and the implications of combining them in different ways can give us a deep understanding of the many *discontinuities* in design space, helping us understand better the differences between different organisms, including the differences between different people (e.g. brain damage may remove or transform a subset of the capabilities). This could provide the basis for new investigations of the evolution of human intelligence, by showing how we correspond to a particular region of design-space, and tracing a succession of regions back to much simpler combinations of capabilities, in other parts of design space. (It isn't generally appreciated that Darwinian evolution, by ruling out inheritance of acquired characteristics, implies that evolution is NOT continuous: there are many small discontinuities between generations.) From this standpoint, arm-chair debates about whether computers or non-human organisms can or cannot ground symbols or use symbols with meaning should be replaced by much deeper investigations into the varieties of architectures and mechanisms that are or are not capable of supporting different sets of combinations of capabilities. Note that the ability to use symbols with different roles is dependent on possession of an architecture with sufficient functional differentiation to support the different causal roles. (This is part of the answer to a common philosophical objection that computers can't use symbols with meaning, because what things mean to you is bound up with what matters to you, and nothing can matter to a computer. This objection ignores the possibility that whether a computer-based system can have desires will depend on the architecture of the system: and very different software architectures can be implemented using the very same underlying computer. Thus, what's true of the computer need not be true of the combination of computer and software.) Motivation for this sort of investigation is undermined by both the assumption of a single dichotomous division and the assumption that there's a continuous spectrum of cases with only arbitrary dividing lines. They are both mistaken: they both ignore the variety of possible designs for behaving systems. Unfortunately people tend to want simple-minded answers to simple-minded questions instead of addressing the rich diversity of reality. In particular, obfuscatory questions like "Is there anybody home?" or "can computers understand?", which assume that there must always be a yes or no answer, divert us from the rich and rewarding philosophical and scientific investigation of design-space, on the basis of which we may better understand ourselves, other organisms and machines of the future. Cheers Aaron From comp.ai.philosophy 5958 Article: 5958 in comp.ai.philosophy Newsgroups: comp.ai,comp.ai.philosophy,comp.ai.nat-lang,sci.lang,sci.cognitive From: A.Sloman@cs.bham.ac.uk (Aaron Sloman) Subject: Re: Origin of "Symbol Grounding Problem"? (Long) Message-ID: Sender: news@cs.bham.ac.uk Organization: School of Computer Science, University of Birmingham, UK References: <256mti$90h@molly.anu.edu.au> Date: Sun, 12 Sep 1993 09:23:10 GMT christo@psych.toronto.edu (Christopher Green) doesn't like my comments on the so-called "symbol grounding" problem: > Organization: Department of Psychology, University of Toronto > Date: Fri, 10 Sep 1993 19:36:27 GMT > > In article I wrote: > [......] > >Many people (e.g. Harnad) hold the mistaken view that it's a single > >all-or-nothing matter whether an organism or machine can "really" > >use symbols to refer to things, or can interpret or understand > >symbols. This tends to be linked with the question whether the > >organism or machine can have experiences. E.g. Harnard assumes we > >can always ask of any machine or organism "Is there anybody home?" > >assuming there must always be a yes or no answer. Christopher responds > This is not the problem, though. I agree that "is there anybody home" is not the problem. In fact all I said was that an all-or-nothing view of the symbol grounding problem "tends to be linked with" an all-or-nothing view of the "is there anybody home" issue. It looks as if Christopher takes an all-or-nothing view at least of symbol grounding. > ...The problem is "How do we *refer* to things > in the world.? Yes. except for the restriction to "we". How do we, other animals and possibly intelligent machines refer to things, is what the problem is about. > No other physical things refers to things apart from > themselves. Well that's an interesting and strong claim. Perhaps a teeny bit dogmatic? Many people think that even bees can refer to locations where they have found food, and communicate what they are referring to by doing a dance which conveys pretty precise information to other bees. If a nest-building bird is away from its nest, collecting materials (twigs feathers etc.) it behaves as if it knows where the nest is, by flying back to it repeatedly. This suggests that there's something in its brain that *refers* to that bit of the world outside it. (I assume you don't believe it happens by magic. If you say that's not a case of reference at all, then I suspect you are committed to the all-or-nothing view, since it is quite widespread.) Many people will claim that their pets can refer to things outside themselves, and can make requests that implicitly involve reference to non-existent states of affairs, e.g. a cat trying to get you to open the door. Then there are those amazing chimpanzees that have been trained to understand a significant subset of English. I saw a TV programme shown on UK television recently, which ended with a woman asking a chimp to do things, including take off her shoes (which it did with great expertise, including undoing the laces when pulling did not suffice). She also asked him to get the X which was outside the door. (X was a ball, or some such thing. I forget what, exactly.) The chimp got up, walked past an X inside the door, fetched the one from outside, and brought it to her. Or perhaps when you say "we" you include other animals? and when you so "no other physical things" you mean apart from humans and other animals? If you really intended to claim that human beings are unique, then you need to produce some argument that all the evidence to the contrary is illusory. The usual way to do that is to fall back on the all-or-nothing dogma, and one way to do that is to say that *real* reference depends on the presence of another all-or-nothing entity (consciousness, spirit, ghost in the machine, or whatever.) But it's not clear what your own reasons are. > How do we do it? Agreed, that's the crucial qustion, where "we" includes all the other animals. Also "which bits of the capability do other animals have that don't have all of our referential capabilities, and how do they do it?" And "which subsets of our referential capabilities might be built into machines of various kinds, and how might this be done". These are questions on which philosophers and engineers need to collaborate: engineers are not skilled enough at making subtle distinctions between different kinds of concepts describing mental capabilities. Philosophers generally lack the experience required to think about possible kinds of mechanisms. (Aaron) > >This is a mistake. Our symbol understanding capability is a complex > >mixture of different sub-capabilities, not all of which need be > >present or absent together. Thus different organisms and machines > >may possess different subsets of these capabilities, with very > >different consequences. (Christopher) > I find it hard to see this as anything more than irrelevant techno-talk. An amazing response. I didn't think I used any words or phrases that could be described as "techno-talk". Also if someone asks about the nature of a capability (which is what the symbol grounding problem, even as described by you, does), then why is it irrelevant to say the capability is composed of many sub-capabilities? You may disagree with the claim, but that doesn't make it irrelevant. > All the "sub-capacities" in the world don't give us an explanation > of how we refer, unless one or more of them does the referring. The fact that four legs keep a table upright doesn't mean that it's impossible for three of them to do something interesting. In fact if you remove one leg, and put a heavy weight over the diagonally opposite corner you can get something close to the original stability. I claim that the ability of human beings to refer is made up of a complex cluster of sub-capabilities (e.g. syntactic capabilities, information gathering capabilities, information storing capabilities, information transforming capabilities, information using capabilities -- e.g. in planning and controlling actions) and that different subsets of these capabilities can occur in other organisms or other machines, which will therefore have capabilities close to but not always necessarily exactly the same as our ability to refer. The papers referred to in my previous posting expand on this in some detail. But I do not claim that the topic has been fully analysed. In particular, we still need a detailed analysis of the functional differentiation in mental capabilities that are support the ability to refer. And we also have only primitive analyses of the semantics of sentential or propositional symbols and practically nothing significant on the semantics of the wide variety of non-propositional forms of representation that seem to be crucial for human capabilities, in both motor control and abstract mathematical reasoning (where non-verbal pictures sometimes play an important role). > And *that's* what we need the explanation of. Increasing the complexity > of the system, in and of itself, does nothing to alleviate > the problem. I wasn't increasing the complexity. I was claiming that the system IS complex, and that we should also consider some of the less complex subsets, such as might occur in other animals. Of course, there are people who say things like "I know what it is to refer because it is what I am doing *now*". I.e. they think they are defining a concept by internal ostensive definition. As a person educated in philosophy you must be aware of the arguments against that sort of definition (e.g. Wittgenstein's attack in his Philosophical Investigations.) Thinking yu can ostensively identify referring, or understanding, is a bit like the mistake of thinking you can specify a point of space by pointing at it, or attending to it, withat realising (as Einstein did) that what exactly you are pointing to depends on which relationshps with other things you are interested in: your finger, the walls of the train you are in, the bit of the countryside through which you are moving, the solar system, etc. I.e. an apparently simple notion like "this point of space" has deep internal complexity, especially as soon as you consider identity over time. Similarly an apparently simple experience of referring has deep internal complexity in that it depends on a host of different capabilities including many of which you are probably quite unaware (like the child that is unaware of the syntactic sophistication she uses in talking English). The all-or-nothing view rejects all this. I've taught philosophy for many years, and I've met that view often. It is not easy to argue against, because it is possible to shore it up with a lot of mutually consistent but in my view equally mistaken philosophical beliefs (e.g. the ghost in the machine type of view that often surfaces in Harnad's discussions.) (Aaron) > >Instead of trying to categorise them using ordinary language or > >traditional philosophical concepts (e.g. "understand", "experience", > >"intentionality", "anybody home"), which are not rich enough for the > > Rich enough in what sense? Intentionality is the name of the > problem, "rich" or no. Yes. We can use ordinary language to start off discussion of the problem. But to find answers we shall have to extend it. Just as people used ordinary language to ask questions about the nature of matter and the universe, but had to extend it (e.g. with the atomic theory of matter) in order to find answers. And when you get a good set of answers you learn that the set of concepts you started with (e.g. the concepts of kinds of stuff -- earth, air, fire, water, iron, etc. don't do justice to the richness of reality -- air can be composed of different combinations of gases at different times, with different consequences for health, etc. water sometimes has hydrogen sometimes deuterium, so there are really at least two kinds, etc....) Why should it be any different with mental concepts? > ..As for "anybody home" this is not a "traditional > philosophical term." It is an off-hand characterization. So what > we're reallyy into here is just philosophy-bashing, right? Why should I bash philosophy? I unashamedly tell people that my main activity is philosophy. I'll bash what I think is bad philosophy, or bad science (like mindless experimentation often found in psychology labs) or bad AI (e.g. building systems without having any clear characterisation of what problem is being addressed), etc. > ...Moreover, > The Really important philosophical terms are "sense" and "reference". > Where do they appear in your list? The literature > on them is as rich as any. Yes I agree that they are important and have written about them. (I regard Frege who made a major contribution in this area, as one of the greatest philosophers of all time.) And a full account of the complex collection of capabilities that I was referring to would explain how reference (what Frege called Bedeutung) normally, but not always, includes sense (something like what Frege called Sinn). But these categories are not rich enough to deal with the full complexity of a working system that can refer: e.g. the human mind or brain. I wrote > > From this standpoint, arm-chair debates about whether computers or > >non-human organisms can or cannot ground symbols or use symbols with > >meaning should be replaced by much deeper investigations into the > >varieties of architectures and mechanisms that are or are not > >capable of supporting different sets of combinations of > >capabilities. Christopher objects: > So it is philosophy-bashing we're doing here: "arm-chair bad, lab good." My words were badly chosen. I used the phrase "arm-chair" not to characterise philosophy, but to characterise the sort of philosophy that people do without finding out a lot more about the rich variety of reality, which goes beyond what they can imagine to be possible. And I also believe that philosophy has now reached the kind of sophistication that makes it impossible to evaluate ideas simply by discussing them. You have to see whether they can be made to work in real designs. My own experience of doing that is that I usually find all sorts of gaps in my ideas in the process of trying to implement them. And when I think I have an implementation, someone comes along and tries a new example that shows that my theories were not sufficiently general. Thus doing philosophy well nowadays, in my view, includes doing AI. That's the main reason I am interested in AI. I suspect Frege, Kant, Aristotle, Leibniz, and many others would have welcomed it for the same sort of reason. > Philosophy has offerred far-deeper discussions of the problems > of intentioanlist and semantics than *any* I've > ever seen in the AI literature. Some AI-ists still think that > SHRDLU was a case of succssful reference (though perhaps on a small scale) > This is grade-school stuff. Agreed. AI whizz kids are often good at marketing their demonstrations in a way that obscures their limitations. Often this is because they have been trained in computing and mathematics but have studied no philosophy, linguistics, psychology, anthropology, etc. (It's not their fault.) (Aaron) > >In particular, obfuscatory questions like "Is there anybody home?" > >or "can computers understand?", which assume that there must always > >be a yes or no answer, divert us from the rich and rewarding > >philosophical and scientific investigation of design-space, on the > >basis of which we may better understand ourselves, other organisms > >and machines of the future. Now Christopher reveals his all-or-nothing philosophy. > Bull. Either systems refer, or they don't. > You can't refer 60% to a chair. Well, maybe a dog or a cat can achieve 60% of our abilities to relate to chairs in our mental processes? Perhaps a chimp can achieve 85% ? Who knows? This is an over simplified response. It needs a lengthy discussion, but this message is already too long. > ...Muddying the waters with > loose talk of "continua" doesn't help a peep. I did not talk about continua. In fact I think that the space of possible designs is mostly discontinuous. We need to understand those discontinuities: a very important task for philosophers of the future. > ..What is > "obfuscatory" is replacing solid debate with technological red-herrings > that "answer" the question by saying that there's is no answer, > or that the wrong question's been asked. Well the "solid debate" that you prefer tends to me to read like a lot of endless assertion and counter assertion. I recommend deeper analysis as a way of making progress. But it's interesting how I've managed, unintentionally, to trigger some kind of technophobia??? Cheers. Aaron -- From Aaron Sat Apr 1 23:49:05 BST 1995 Newsgroups: comp.ai.philosophy,comp.ai,comp.robotics,comp.cog-eng,sci.cognitive,sci.psychology Summary: "grounding" is the wrong concept References: Subject: Re: Grounding Representations: ("Grounding" is the wrong word) departed@netcom.com (just passing through) writes: > Date: Fri, 24 Mar 1995 02:33:46 GMT > > In article , > Robert White wrote: > >In harnad@ecs.soton.ac.uk (Stevan Harnad) writes: > > > >[.] > >>Intelligence is that computer programs use symbols that are arbitrarily > >>interpretable (see Searle, 1980 for the Chinese Room and Harnad, 1990 > >>for the symbol grounding problem). We could, for example, use the word > >>"apple" to mean anything from a "common fruit" to a "pig's nose". All > >>the computer knows is the relationship between this symbol and the > >>others that we have given it. > > > > > >Systems theory provides a grounded approach to solving this problem > >and I have seen the same 'signification' models used within > >metamodeling as I have within Semiotics. I was especially surprised to > >see almost the exact same model used by Roland Barthes in his book > >entitled Mythologies. The structure of the model is tripartite and > >each signal is generated to create a 'signifier' and a 'signified' > >semiotic meaning. Note that any theory of meaning that requires there to be an existing object that is referred to (a `signified') is just WRONG as a theory of how human-like intelligence works. We frequently refer to things that don't exist (Mr Pickwick, Unicorns, that parrot sitting on your left shoulder, imaginary scapegoats, dieties, expected disasters that don't materialise, the largest prime number between 24 and 28, and many other things.) We also refer to things that are inaccessible in space and time and to things about whose existence we are unsure (Was he murdered, and if so by whom?). These are not just foolish quirks and foibles: the ability to create meaning without assuming a referent (discussed at length a century ago by Gottlob Frege, from whom we also got predicate logic, higher order functions, and indirectly lambda calculus, Lisp and Prolog) is essential for forming goals (which may not be attainable), for making plans (including plans for contingencies that may not arise), and for asking questions which drive the search for knowledge and understanding (and therefore power). It's also a consequence of having inaccurate, or out of date information. (Actually Frege, having identified the possibility of meaning without a referent made the bizarre proposal that a special referent "The False" be associated with things that don't refer.) Anyhow, because meaning does not require a referent, I claim that Harnad has (unwittingly) done much harm to AI and cognitive science by introducing the term "symbol grounding" to express a problem that was well understood previously (the problem of explaining intentionality in humans or machines: previously discussed by Hume, Husserl(?), Russell, Wittgenstein, Haugeland, Searle, Dennett, and many others.) The phrase "symbol grounding problem" misleadingly suggests that every meaningful symbol has to be "grounded", and that leads to misguided theories requiring meaning to arise out of some sort of "contact" or fairly direct causal connection with reality, like the misguided "robot reply" to John Searle's chinese room argument. I think it is more accurate to regard meaning as arising primarily out of structure and internal manipulations. External causal links are needed to reduce residual ambiguity, but can never remove it completely. (Without external links a robot might use the concept of (something like) a tower or a poet, but not could not have a concept of THE Eiffel Tower, or of William Shakespeare.) Most of our deep theories about the world use concepts that are not "grounded" in perception or causal connections. (The requirement for such grounding is just a re-incarnation of the concept empiricism of old philosophers like Berkeley and Hume, which was demolished by Kant, who argued that in order to acquire concepts from experience you'd have to have some concepts to start with, in order to have experience.) Some of this stuff is discussed at length in many text books on philosophy of science, for science is full of concepts not grounded in causal linkages, e.g gene, quark, neutrino, electromagnetic field, natural selection. It's a long story. I have two papers elaborating on this: `What enables a machine to understand?' in Proceedings 9th International Joint Conference on AI, pp 995-1001, Los Angeles, August 1985. `Reference without causal links' in Proceedings 7th European Conference on Artificial Intelligence, Brighton, July 1986. Re-printed in J.B.H. du Boulay, D.Hogg, L.Steels (eds) Advances in Artificial Intelligence - II North Holland, 369-381, 1987. Both are available in the Cognition and Affect ftp directory: ftp://ftp.cs.bham.ac.uk/pub/dist/cog_affect The files are Sloman.ecai86.ps.Z Sloman.ijcai85.ps.Z I make heavy use of Rudolf Carnap's concept of a "meaning postulate", explained in his book Carnap, R., Meaning and Necessity Phoenix Books 1956. There's still much work to be done! Aaron --- From A.Sloman Sun Apr 30 17:52:13 BST 1995 Newsgroups: comp.ai.philosophy,comp.ai,comp.robotics,comp.cog-eng,sci.cognitive,sci.psychology References: <3lkl8d$2gm@percy.cs.bham.ac.uk> <3lkrpq$kun@mp.cs.niu.edu> <3nhlk5$i7o@percy.cs.bham.ac.uk> Message-ID: <3o0f74$l97@percy.cs.bham.ac.uk> Subject: Re: Grounding Representations: ("Grounding" is the wrong word) Some replies to interesting comments from Andrzej Pindor (pindor@gpu.utcc.utoronto.ca - 27th April) and Mark Rosenfelder (markrose@spss.com 25th April), and Oliver Sparrow (ohgs@chatham.demon.co.uk 27 April) [Apologies to people whose comments I've missed.] Andrzej Pindor writes: > In article <3nhlk5$i7o@percy.cs.bham.ac.uk>, > Aaron Sloman wrote: > ......... > >In a thermostat, and perhaps some simple organisms, the link between > >internal information store and and thing represented is very > >direct. In people the causal links are very indirect and the same > >sensors (and motors) are shared between huge numbers of different > >concepts, with many intermediate levels of processing between > >sensory transducers and states like beliefs. > > > >The more indirect (and overloaded) the causal links between > >representations and referents, the more the meaning depends on > >structure not causation. In humans I believe structure dominates, > >and causal links serve merely to reduce ambiguity of reference > >(which can never be completely eliminated). > > > >The structure of our internal information states is so rich, and the > >architecture that uses them is so complex that the bulk of human > >meaning comes from the interaction of structure and manipulation. Andrzej responded > Experiments with very young kittens, who from birth were brought up in an > environment with vertical lines only and which were found later to be unable > to see horizontal lines seem to suggest very strongly to me that a large ^^^^^^^^ > part of what you call structure and manipulation has its source in causal ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ > links. ^^^^^^ > ............. The conclusion is surely correct, whatever kittens may do in such experiments. To engage the world successfully a robot, or kitten, cannot (normally) create its internal model of the world entirely on the basis of apriori hypotheses (or prejudices). The global process of building up the internal representations that play a role in taking decisions, planning, interpreting subsequent sensory input, etc. must, at least in part, be shaped by the agent's interactions with the environment, for otherwise it would be pure hallucination. (Maybe for some pour souls it is: they would then need a lot of help from others in order to avoid doing themselves harm.) [Let's for now discount lucky guesses.] Oliver Sparrow (ohgs@chatham.demon.co.uk) seemed to be saying the same sort of thing, when he wrote (27 Apr): > The grounding issue seems to be acute when one takes a snapshot of a > situation and tries to assess it in terms of structures which are > "in here" versus "out there". The 'impossibility of AI because of > the grounding problem' then does indeed seem to be a genuine > problem. > > Actually, of course, all that we know about natural intelligence > points to *process* as being key. Systems iterate towards internal > descriptions of structures of which they are not directly a part. > The correspondence I am not denying the existence or usefulness of causal links. It's the relative importance of such links vs internal structure and mechanism that I say varies according to the kind of agent. The more diverse the agent's store of information the less its semantics can depend on the sorts of causal links that you get in simple control systems. Note that there are two kinds of relevance of causal links: (1) accounting for the origins of the stored information (2) accounting for its current semantic content. In both cases the relative importance of causal links will be less, the more sophisticated the agent's cognitive capabilities. (I am, of course, using vague concepts for which there are no agreed measures: e.g. relative importance, sophistication.) So I don't dispute the relevance of causal connections. It's blindingly obvious, especially as regards (1)! But exactly HOW causal links are relevant to (2), and what else is needed, is not so obvious. Immanuel Kant, who argued strongly that intelligent agents require a great deal of non-trivial apriori knowledge, argued something like this: Even if all our actual knowledge is triggered by having sensory experiences, it does not follow that the knowledge is all derived from those sensory experiences. (I don't have my copy to hand, but I think this was in the Introduction to the second edition of "Critique of Pure Reason". I think he used words like `awakened' rather than `triggered' and `arises out of' rather than `derived from'. But my memory is hazy, and the translation I read could have been inaccurate.) I see one of the still unfinished tasks of AI as being to sort out exactly what was right and what was wrong about what Kant wrote. Mark Rosenfelder, took a similar line to Andrzej. In on Tue, 25 Apr 1995 18:04:57 GMT, Mark wrote: > But even if you were right, and the "structures" overwhelm the > "experience" in complexity, even that does not dispose of grounding. Note that if the word "grounding" is interpreted in a sufficiently general and loose way, I am not opposed to grounding of meaning. I was objecting to the narrow and simple-minded sort of direct link that the word invites the philosophically inexperienced to assume is required. Even truth can be harmful if expressed the wrong way. Mark continues: > The word "dog" may be linked to huge masses of purely conceptual > information, from reading, talking, or reasoning; but it's still > linked to actual experiences with dogs (and with other animals, with > fur, with our own bodies, etc.-- the experiences that lend meaning > for me to words like "llama", although I've only seen real llamas a > few times). A lot depends on how direct you think those links have to be, and whether you think they have to be YOUR experiences, and whether you think the links have to continue to operate in order to sustain semantic content, and whether you think the causal links somehow provide a SUFFICIENT basis for the concepts, or whether, like me, you regard them as merely a necessary part of a larger system of relationships, e.g. playing the role of reducing ambiguity, in the sense described below. It's important to separate out several questions: (a) does the ability to think about things in the environment depend on the agent having causal connections with the environment. My answer was clearly yes, though I claimed that because the causal links are so weak and indirect in human-like intelligent agents (as opposed to thermostats) those links cannot do the job of identifying particular referents. (Maybe house-flies and many other organisms are closer to thermostats. Kittens may be closer to humans.) Note that this is not a claim about the previous causal links or future possible links. I am talking about what makes it possible for you NOW to think about Julius Caeser, the square root of 1000, what you are going to have for dinner tomorrow, whether there will be peace on earth by the year 3000, why dinosaurs became extinct, etc. etc. My claim is that existing causal links play very little role in determining the semantic content of most of the information stored in your brain (or mind) right now. At this moment there's a huge amount of information about your immediate environment, about things remote in time and space, about generalisations, legal rules, family relationships the grammar of the language(s) talked in your culture, etc. The causal links between the fine detail of all that information and the corresponding bits of the environment are either very weak or effectively non-existent at the moment, except for those aspects of your internal state that relate to your immediately perceived (and acted on) current environment. E.g. most of the other bits of the world could change in some way without affecting your representations of those bits, and vice versa. That's why I claim that the mechanisms and the structure of representing states are more important than their causal links as a basis for your ability (at any given time) to think about their referents. Structure turned out to be more useful than the causal links because the structure of a representing state is more enduring and more accessible to the rest of the system for purposes of reasoning, planning, asking questions, etc. The ability to compile information into accessible enduring structure is central to intelligence. I claim that after this ability evolved it increasingly dominated information processing (in the relevant organisms), as contrasted with "online" control mechanisms where causation dominates in determining semantic content. (I'll challenge the requirement for historical causal links below.) As a first approximation to what I am getting at, consider Tarski's work on truth. He gave a recursive definition of the conditions under which a set of objects (with various properties and relations) could be a model for a set of axioms expressed in predicate calculus. That's an example of what I mean by semantics depending on structure. (NB the model and the axioms need not be isomorphic: e.g. a small number of axioms could have an infinite model, like Peano's axioms for arithmetic.) Tarski's ideas would need to be generalised to extend to other forms of representation than predicate logic, e.g. pictures, computer programs, neural representations, etc. However, Tarskian semantics *obviously* cannot identify a *unique* model for a given set of representations. For, if any M is a (tarskian) model for S, and M' is isomorphic with M, then M' is also [Previous line corrected since original posting] a (tarskian) model for S even if M' is millions of light years away in some other galaxy. That's where causal links can come in. If S is embedded in a mechanism providing a web of causal links with the environment, then that can (sometimes) be a basis for eliminating M' as referent Alas, the combination of structure and causation cannot *totally* eliminate ambiguity. In the late 1950s and early 60s (when I was a philosophy research student) philosophers used to talk about "open texture" of language (I think the term was introduced by Friedrich Waismann, perhaps under the influence of Wittgenstein). Alas, I again don't have a reference to hand. (Perhaps it was in his articles on analytic-synthetic in Analysis, circa 1949). I claim that not only language, but also thought, perceptual states, desires, etc. are all (to some extent) open textured in that they do not totally unambiguously, uniquely refer to things. This open texture, far from being a flaw, provides growth points for new concepts, allowing us to go on gradually extending our understanding of the universe without constantly having to introduce radical revision. (But that's another long story.) Returning to varieties of questions about causal vs structural bases for semantic lnks: (b) does the semantic relation between some internal state S and some object O in the environment that it refers to depend on permanent causal links between S and O? Answer NO, for the reasons indicated. (We have two few causal channels to dedicate them to preserving such long term correspondences. Imagine being stuck forever staring at the Eiffel Tower to make sure that your internal representation changed if it did.) (c) is there any particular sort of causal link that is uniquely suited as a basis for helping to pin down semantic content? Answer NO. Some people have claimed that a system whose interaction with the environment was restricted to the use of a character terminal to give instructions to other agents and which other agents use to feed back information, could Not accurately be said to understand a conversation about the environment. For instance it might be claimed (and I think Harnad does claim) that the causal interaction MUST go via analog sensory transducers that feed directly in to neural nets, and via motors (muscles) that are driven directly by control signals coming out of neural nets. I have never seen a convincing argument in support of such restrictions in the causal interactions capable of sustaining semantic links. On the other hand the theory according to which the main role for causal links is to pin down a model and reduce ambiguity need not care exactly what kind of causation it is, provided that it eliminates the right models. On this theory, it is not possible for my thinking about (e.g.) the Eiffel Tower to be about that particular tower PURELY in virtue of internal states and processes. My causal links to the tower are also needed, in order to rule out other objects that might be models for my thoughts, beliefs, etc. But the links can be very indirect, and may depend only on my being located in space and time in such a way that I can relate to that object via other people, possible forms of transport, news broadcasts, etc. These external relationships suffice to prevent me unwittingly referring to some object just like the Eiffel tower in another part of the Universe. (The need for such causal embedding to remove ambiguity is one of the reasons why merely having some super NLP program running inside you cannot be a SUFFICIENT basis for understanding English. But I don't regard that as an attack on *sensible* variants of Strong AI) (c.1) Must any of the concepts I use have come via use of particular sorts of sensors? NO There are more detailed issues about exactly which concepts you can or cannot understand about the environment if you lack specific sorts of sensors. People have lots of prejudices about this because we all tend to be attracted to concept empiricism, the view which says you cannot talk about something unless you have experienced it directly yourself. If true, this would make it impossible for evolution to drive the development of organisms that are born with rich knowledge about the world: e.g. a new-born deer can run with the herd within hours of its birth, without having to LEARN how to interpret all those photons falling on its retina as representing a 3-D environment requiring particular motor processes to navigate it. Of course, its ancestors' interactions with the environment played a role in giving it this capability. Similarly, congenitally blind humans may have much of the apparatus required for understanding talk about colours, how things look etc. Their understanding is not directly rooted in THEIR experience, but in an evolutionary history involving visual contact with spatial phenomena. (c.2) Must the concepts have come via use of particular sorts of sensors either in the agent or in its evolutionary forebears? NO If by a highly improbable fluke of mutation an animal were born with the visual and other capabilities of a young deer WITHOUT this being the result of previous selective pressures, that would not mean the new sport could not see or have intentions relating to the environment as well as a new foal: the CURRENT internal structures and mechanisms, and causal links would suffice, for all practical purposes, without the normal causal history. (Note that I am not arguing that this is likely to occur: merely that there's nothing logically impossible about it. Similar points are made by Roger Young, in The Mentality of Robots, {\em Proceedings Aristotelian Soc. 1994}) (Of course, this example will not stop prejudiced philosophers from saying: "this animal does not `Really' see, or think, or take decisions, despite appearances, because it does not have the right evolutionary history"! But then we get involved in disputes about essentially trivial matters of definition. I'll can define two notions of "see": one capability (to seeH) requires a normal historical source, and the other (to seeA) is a-historical. Apart from that there is no difference in the details of the capabilities, i.e. how well they enable the organism to survive. I then have a ready made way of talking usefully about the new specimen whose ability to seeA is not rooted in evolutionary history. Those who insist on using only "seeH" will find it very cumbersome to describe the same animal.) (Aaron) > >(b) In fact it may turn out easier to design and implement a > >disembodied (or perhaps I should say "disconnected") mathematician > >whose mind is concerned with nothing but problems in number theory > >(and who enjoys the thrill of discovery and experiences the sorrow > >of refutation) than it is to design and implement a robot with > >properly functioning eyes, ears, arms, legs, etc. > > Andrzej commented: > In case of abstract mathematical terms it may very well be that their > complete meaning is contained in a web of internal links, with no causal ^^^^^^^^^^^^^^ > links involved. In fact Harnad, asked whether one can talk about meaning ^^^^^^^^^^^^^^ > of abstract mathematical terms in view of his concepts on grounding, > ducks the question. On the other hand it is not unlikely that when we ^^^^^^^^^^^^ > think about abstract mathematics, we do so by mapping mathematical > terms and their relationships onto mental structures which have come > into being by our exposition to sensory stimuli. Note that I am not concerned with what is or is not likely or unlikely in the case of human beings, but with what is theoretically or technically possible, e.g. for artificial agents. Neither did I say that NO causal links would be required for the disconnected mathematician. On the contrary, in order to have the thrill of discovery and the sorrow of failure the mathematician will need a rich internal architecture with lots of causal links between internal subsystems. The internal mechanisms that operate on internal structures will involve many intricate causal links. (See my paper in IJCAI 1985). Motivational and emotional states require especially rich mechanisms with internal causal links. But that's another whole story. For discussions on architectures underlying motivational and emotional states see my papers, and papers by Wright and Beaudoin, in ftp://ftp.cs.bham.ac.uk/pub/dist/cog_affect (aaron) > >Of course, if the mathematician really lacks sensors and motors, > >then we shall have no way of finding out which theorems it is > >exploring etc., unless we can use our knowledge of its design and > >direct measurement of internal physical states. But this will be > >analogous to decompiling a machine code program, which can be > >impossibly difficult. > > > >Anyhow the important thing is not to speculate about what is > >possible, but to get on and do it, or find out exactly why it is > >impossible. So let's have a go at designing the mathematician. [my spelling corrected] (Andrzej) > Such 'mathematicians' are being designed. A programm Graffiti by Siemion > Fajtlowicz from University of Huston may be a case in point. An interesting > thing is that such programms work differently than a human mathematician > (for instance they have no 'mental structure' derived from sensimotoric > stimuli, suggested above) and hence may work out results (conjectures in case > of Graffiti) which would not occur to a human. I don't know this work. Similar work is being done (I think) by Edmund Furse at the University of Glamorgan (his program learns mathematics by being presented with the contents of university level text books on e.g. group theory, expressed in a suitable formal language. It develops the ability to solve the exercises in the text book. There is also a lot of work on automatic theorem proving, etc. I don't think anyone working in that sort of area shares my interest in trying to understand the architectural basis for motivation, emotions and other affective states. I suspect it will be a long time before an artificial mathematician gets excited about a theorem it has proved. On the other hand, the sort of work already going on may contribute to its design. > Andrzej Pindor The foolish reject what they see and > University of Toronto not what they think; the wise reject > Instructional and Research Computing what they think and not what they see. > pindor@gpu.utcc.utoronto.ca Huang Po The really wise weigh up both what they see and what they think and try to optimise the relationship by rejecting either, as necessary. [Sorry to go on so long. I am trying to write a paper on all this stuff.] Aaron --- From Aaron Sun Apr 30 19:25:38 BST 1995 Newsgroups: comp.ai.philosophy,comp.ai,comp.robotics,comp.cog-eng,sci.cognitive,sci.psychology References: <3lkl8d$2gm@percy.cs.bham.ac.uk> <3lkrpq$kun@mp.cs.niu.edu> <3nhlk5$i7o@percy.cs.bham.ac.uk> <3o0f74$l97@percy.cs.bham.ac.uk> Subject: Re: Grounding Representations: ("Grounding" wrong word) [CORRECTION] Alas, I wrote > However, Tarskian semantics *obviously* cannot identify a *unique* > model for a given set of representations. For, if any M is a > (tarskian) model for S, and M' is isomorphic with S, then M' is also ^^^^^^^^^^^^^^^^^^^^^^^ > a (tarskian) model for S even if M' is millions of light years away > in some other galaxy. That's where causal links can come in. If S is > embedded in a mechanism providing a web of causal links with the > environment, then that can (sometimes) be a basis for eliminating > M' as referent Instead of the underlined bit, I meant to wrote M' is isomorphic with M Isomorophism between M' (a model) and S a set of axioms, is irrelevant. Sorry if this generates any confusion. Aaron [more discussion followed]