Date: Fri 12 Aug 1988 15:35-EDT From: AIList Moderator Nick Papadakis Reply-To: AIList@mc.lcs.mit.edu Us-Mail: MIT LCS, 545 Tech Square, Rm# NE43-504, Cambridge MA 02139 Phone: (617) 253-6524 Subject: AIList Digest V8 #47 To: AIList@mc.lcs.mit.edu Status: R AIList Digest Saturday, 13 Aug 1988 Volume 8 : Issue 47 Philosophy: AI and the future of the society Dual encoding, propostional memory and... Self-reference in Natural Language point of metalanguage in language The Godless assumption ---------------------------------------------------------------------- Date: 5 Aug 88 17:24:30 GMT From: jbn@glacier.stanford.edu (John B. Nagle) Reply-to: glacier!jbn@labrea.stanford.edu (John B. Nagle) Subject: Re: AI and the future of the society Antti Ylikoski (YLIKOSKI@FINFUN.BITNET) writes: >I once heard an (excellent) talk by a person working with Symbolics. >(His name is Jim Spoerl.) > >One line by him especially remained in my mind: > >"What we can do, and animals cannot, is to process symbols. >(Efficiently.)" > > >In the human brain, there is a very complicated real-time symbol >processing activity going on, and the science of Artificial >Intelligence is in the process of getting to know and to model this >activity. > >A very typical example of the human real-time symbol processing is >what happens when a person drives a car. Sensory input is analyzed >and symbols are formed of it: a traffic sign; a car driving in the >same direction and passing; the speed being 50 mph. There is some >theory building going on: that black car is in the fast lane and >drives, I guess, some 10 mph faster than me, therefore I think it's >going to pass me after about half a minute. To a certain extent, the >driver's behaviour is rule-based: there is for example a rule saying >that whenever you see a red traffic light in front of you you have to >stop the car. (I remember someone said in AIList some time ago that >rule-based systems are "synthetic", not similar to human information >processing. I disagree.) As someone who works on automatic driving and robot navigation, I have to question this. One notable fact is that animals are quite good at running around without bumping into things. Horses are capable of running with the herd over rough terrain within hours of birth. ("Horses of the Camargue" has some beautiful pictures of this.) This leads one to suspect that the primary mechanisms are not based on symbols or rules. Definitely, learning is not required. Horses are born with the systems for walking, obstacle avoidance, running, standing up, motion vision, foot placement, and small-obstacle jumping fully functional. More likely, the basics of navigation are based on geometric processing, or what some people like to call "spatial reasoning". See Witkin and Kass's work at Schlumberger for some idea of what this means. Ossana Khatib's approach to path planning (1979) is also very relevant. Geometry has the advantage of being compatible with the real world without abstraction. Recognize that abstraction is not free. In real-world situations, as faced by robots, the processing necessary to put the sensory data into a form where rule-based approaches can even begin to operate is formidable, and in most non-trivial cases is beyond the state of the art. I would encourage people moving into the AI field to work in the vision, spatial, and geometric domains. There are many problems that need to be solved, and enough computational power is becoming available to address them. Much of the impetus for the past concentration on highly abstract domains came from the need to find problems that could be addressed with modest computational resources. This is much less of a problem today. We are beginning to have adequate tools. Personally, I suspect that horse-level performance in navigation and squirrel-level performance in manipulation can be achieved without any component of the system using mathematical logic. John Nagle ------------------------------ Date: 5 Aug 88 10:52 PDT From: hayes.pa@Xerox.COM Subject: Re: Dual encoding, propostional memory and... > (yes, I encode all my knowledge of a scene into >little FOPC like tuples, honest) ...... >... thinking about bit level encoding > protocols, .... No: thats not the claim. You are here talking about the implementation level of encoding: we are talking about the semantic level. You might encode your epistmologically adequate ( see McCarthy and Hayes 1969 ) representation in all sorts of ways, perhaps as states of a connectionist network ( although I havnt yet seen a way in which it could really be done ), probably not as lots of little n-tuples ( very inefficient ). The point at issue was whether the knowledge is encoded or not, and whether, if it is, we can make much progress without thinking about how it does its representing. >The dual coding theory, which normally distinguishes >between iconic and semantic memory, has caused >endless debate Yes, but much of this debate has been between psychologists, and so has little relevance to the issues we are discussing here. These trails arent in different directions to McCarthys, they are in a different landscape. Ive had interesting arguments with some of them, about such terms as `iconic memory'. Are there iconic and propositional representations in the head? Of course, the psychologist says; if we can visualise, producing different behaviour than when we remember: thats what `different' MEANS. Thats not what the AI modeller means by `different', though. If one takes the iconic/semantic distinction to refer to different ways in which information can be encoded, then it isnt at all obvious that different behavior means different representations ( though it certainly suggests different implementations ). >Pat's argument hinges on the demand that we think >about something called representation (eh?) and then >describe the encoding. The minute you are tricked... Well now, lets be clear. The argument goes like this. People know things - facts, lets say, but use a different word if you like - and their behavior is influenced in important ways by the things they know and what they are able to conclude from the. It seems reasonable to conclude that these facts that they know are somehow encoded in their heads, ie a change of knowledge-state is a change of physical state. Thats all the trickery involved in talking about `representation', or being concerned with how knowledge is encoded. All the rest is just science: guesses about how this encoding is done, observations about good and bad ways to describe it, etc.. Do you disagree with any of this, Gilbert? If so, what alternative account would you suggest for describing, for example, whatever it is that we are doing sending these messages to one another? >PDP networks will work of course,... Well, will they? Lets see them do some cognitive task of the sort McCarthy has been aiming at. But it must be possible, I agree, to implement it all this way, since we are ourselves walking, talking networks. > ....but you can't of >course IMAGINE the contents of the network, and thus >they cannot be a representation Sure they can be a(n implemetation of a ) representation. And sure we can imagine the contents of the network: people do it all the time. When someone shows me a network doing a bit of semantic memorising, you can bet they are explaining to me how to imagine whats in the network. If people who attack AI or the Computational Paradigm, simultaneously tell me that PDP networks are the answer, I know they havnt understood the point of the representational idea. Go back and (re)read that old 1969 paper CAREFULLY, Gilbert, before you find yourself writing a book like Dreyfusss. Pat Hayes ------------------------------ Date: Tue, 9 Aug 88 10:55 CDT From: Subject: Self-reference in Natural Language In AIList Digest vol. 8 num. 29 I claimed that Bruce Nevin's (bnevin@cch.bbn.com) analysis of self-reference in natural language entailed there was something semantically improper about sentences like "This sentence is in English" and "This sentence is grammatical." My claim was that they are wholly unproblematic, and hence that there was something wrong with the analysis. In his lengthy and very interesting reply, Nevin demurs: > They are not "wholly unproblematical," they engender a double-take kind of > reaction. Of course people can cope with paradox, I am merely accounting > for the source of the paradox. First, I'm dubious about whether they do engender the sort of double-take Nevin refers to here. But even so, it's not at all clear to me what that's supposed to signal. People sometimes have a similar reaction to sentences containing several negatives, but for all that we wouldn't want to cast aspersions on them. Second, Nevin seems to be implying that people have to "cope with paradox" when they are confronted with the self-referential sentences above. But again, even if we grant that there are problematic, there's surely no paradox lurking anywhere nearby. > If I say it in Modern Greek, where the noun followed by deictic can > come last, the normal reading is still for "this" to refer to a nearby > prior sentence in the discourse. The paradoxical reading has to be > forced by isolating the sentence, usually in a discourse context like > "The sentence /psema ine i frasi afti/, translated literally 'Falsehood > it is the sentence this', is paradoxical because if I suppose that it > is false, then it is truthful, and if I suppose it is truthful, then it > is false." These are metalanguage statements about the sentence. The > crux of the matter (which word order in English only makes easier to > see), is that a sentence (or any utterance) cannot be a metalanguage > statement about itself--cannot be at the same time a sentence in the > object language (English or Greek) and in the metalanguage (the > metalanguage that is a sublanguage of English or of Greek). The assumption that there is a metalanguage/object language distinction in ordinary language is carrying an awful lot of weight here. There's no doubt we have to make and heed such a distinction when we're doing formal semantics--where we've actually got a rigorously defined formal language, and we're describing how it's to be interpreted in a formal model--but it's not clear there is such a distinction to be made in natural language. It seems to me far more natural just to say that English (for example), in addition to containing terms that refer to planets, numbers, and the like, also has terms that refer to elements of the language itself, and in the limiting case, to expressions that contain those very terms. Granted, this is precisely the capacity that leads us to paradox; but I would like to see some evidence that this account is wrong other than the fact that it gets us into trouble. That is, is there any linguistic intuition that Nevin can cite to justify his account in addition to his claim that he's got a solution to the paradoxes? Consider an analogy in set theory. Russell's paradox initially engendered all sorts of confusion and consternation--the intuitive assumption that for every property there is a corresponding set of things that have the property leads to contradiction. Eventually, there came something of an explanation in the form of the so-called {\it iterative} conception of set--sets are collections that are "built up" from some initial collection of atomic elements by certain operations. Some properties pick out collections that could never be the result of any such building-up process, and hence are not SETS. Not exactly airtight, but there is an appealing intuition there. Is there anything analogous for Nevin's account? I'm not being rhetorical here; there may well be, I just haven't been able to think of any. Chris Menzel cmenzel@tamlsr.bitnet chris.menzel@lsr.tamu.edu ps: In my first reply to Nevin I included two books for "Recommended reading." This I fear came off looking as if I were patronizing him, which was most definitely not my intent, and I apologize for the misimpression. I have learned a great deal from both books (Martin's {\it Recent Essays on Truth and the Liar Paradox} and Barwise and Etchemendy's {\it The Liar: An Essay on Truth and Circularity}), and my recommendations were sincere, and intended as a genuine contribution for the benefit of the readers of AIList Digest. ------------------------------ Date: Tue, 9 Aug 88 16:07:24 EDT From: "Bruce E. Nevin" Subject: point of metalanguage in language We are talking about self-referential sentences of the type: This sentence is English { grammatical | long | . . . } I agree that these sentences are paradoxical only if the adjective is something like `false, a lie'. You get paradox only when successive readings contradict each other and must be reconciled because it is after all but one sentence. The contradiction makes it impossible to ignore the semantic problem of not being able to resolve referentials. In the paradoxical case, the infinite regress of reading-tokens cannot be ignored because of the contradiction. But even without the contradiction between successive readings, you always get an effect that you might call `referential reverbration', since to evaluate the truth or appropropriateness of a sentence containing a deictic you naturally examine the thing to which it refers, which in these cases happens to be the sentence itself. The rereading (checking out the referent) refers again to itself. Most language users don't continue doing this for very many iterations. Stopping runaway loops presumably has some adaptive value for intelligent entities! Contrast the following case: Cesuwi tini:maCQati. The preceding sentence is in Achumawi. No reverbration, just one glance back. (Uppercase letters are for glottalized stops.) I think self-referential sentences (sentences that refer to themselves as a whole, not just to words or constructions in themselves) are perhaps initially amusing and then later annoying to people because they are rather a perversion of the machinery of deixis. Such sentences occur only in the most artificial circumstances. Virtually anything they can say about themselves is self evident and therefore redundant. Except for illustrating some oddities about language, they are pointless, whether true or false. I mean, who cares that `Afti i frasi ine st anglika' or `This sentence is in Greek' is false? To say such a self-evident falsehood is foolishness without even the point of humor in any circumstances that I can think of. This surely contributes to the feeling of anomaly and is another reason why they are not `wholly unproblematic,' but the real tale is in the process of resolving referentials, as described previously. There is a somewhat similar case in which simple falsehood goes usually unnoticed, an oversight which is certainly not characteristic of non-selfreferential situations: I am thinking here of the familiar notice on an otherwise blank page that says `this page intentionally left blank'. This is like +-------------------+ + This box is empty + +-------------------+. Is the preceding a sentence in English? It would not go unnoticed. One can construct cases like This sentence, about `psemata sta anglika,' is in English. Afti i frasi, epi `falsehoods in Greek,' sta elinika ine. but they too would scarcely go unnoticed. Hofstedter plays these recursion games very nicely. Multiple negatives do indeed involve similar processes, since negation is a metalinguistic operator, one of denial. Such cases are problematic because language users try to resolve them to a simple assertion or denial. (Easier to do in languages or dialects admitting multiple negation for intensive expression, as in West African languages and Black English.) Things like `I don't disagree' and `not unlike the denial of' are considered stylistically bad but are not semantically flawed in the way that self-referential sentences are, because the multiple rereadings have a limit rather than implying infinite regress. One might argue that inability to resolve referentials is a matter of performance rather than competence. That limb is open if anyone wants to go out on it. The point about natural language containing its own metalanguage is not that it can be abused in degenerate cases, but rather that we need no other a prioristic metalanguage for grammar and semantics. Indeed, language users have recourse to no such external metalanguage for learning and using their native language. If a description of a language cannot be stated in the language itself (that is, in its intrinsic, built-in metalanguage), then it is incorrect. It is bound to introduce redundancy into the description over and above the redundancy used by the language for informational purposes, and this has the status of noise obscuring any account of the information in texts. Formal notations may be convenient for computational and other purposes, but they must be straightforward graphic variants of words and constructions in the metalanguage that is a part of the language that they describe. The question of the status of the metalanguage thus points to a criterion for comparison of different descriptions as to their adequacy for representing language and what it does. See e.g. Z. S. Harris _Language and Information_. My review of this book should appear in _Computational Linguistics_ 14.4, scheduled to be mailed in January 1989. Bruce Nevin bn@cch.bbn.com PS: I have not read the books you cite but will look for them. ------------------------------ Date: Thu, 11 Aug 88 11:58:35 GMT From: IT21%SYSB.SALFORD.AC.UK@MITVMA.MIT.EDU Subject: The Godless assumption Subject: The Godless assumption In going through my backlog of AI Mail I found two rather careless statements. In article <445@proxftl.UUCP>, bill@proxftl.UUCP (T. William Wells) writes: > that,.... This means > that I can test the validity of my definition of free will by > normal scientific means and thus takes the problem of free will > out of the religious and into the practical. Why should 'religious' not also be 'practical'? Many people - especially ordinary people, not AI researchers - would claim their 'religion' is immensely 'practical'. I suggest the two things are not opposed. It may be that many correspondents *assume* that religion is a total falsity or irrelevance, but this assumption has not been proved correct, and many people find strong empirical evidence otherwise. Date: Sun, 03 Jul 88 03:47:51 EST Jeff Coggshall writes >Subject: Metaepistemology & Phil. of Science > Once we assume that there is no priveledged source knowledge about >the way things really are, then, it seems, we are left with either >saying that "anything goes" ... That there is 'no priviledged source knowledge' is a mere assumption that has very little evidence to support it. And there are many who do not make that assumption, believing in religious revalation. Many would claim the Bible, for instance, is God's revelation to humankind. Some other religions would make equivalent claims. Therefore we cannot *assume* such things without the danger of writing off a huge section of reality which our theories should fit. Since the non-existence/irrelevance of God has not yet been proved, and many claim to have strong empirical evidence of God's existence and effectiveness in their lives, may I ask that correspondents think more carefully before making statements like the two above. Thank you, Andrew Basden. ------------------------------ End of AIList Digest ********************