gilbert@cs.glasgow.ac.uk (Gilbert Cockton) (06/01/88)
In article <1792@pt.cs.cmu.edu> kck@g.gp.cs.cmu.edu (Karl Kluge) writes: >> Because so little of our effective knowledge is formalised, we learn >> in social contexts, not from books. I presume AI is full of relative >> loners who have learnt more of what they publicly interact with from >> books rather than from people. > >You presume an awful lot. Comments like that show the intellectual level >of your critique of AI. I also presume that comparative attitudes to book and social knowledge are a measurable, and probably fairly stable, feature of someone's make up. It would be intriguing to test the hypothesis that AI researchers place more faith in the ability of text (including programs) to capture social reality than other academic groups. Now, does this still have no intellectual respectability? > >Well, what kind of AI research are you looking to judge? If you're looking >at something like SOAR or ACT*, which claim to be computational models of >human intelligence, then comparisons of the performance of the architecture >with data on human performance in given task domains can be (and are) made. You obviously have missed my comments about work by John Anderson and other psychological research. If AI were all conducted this way, there would be less to object about. >If you are looking at research which attempts to perform tasks we usually >think of as requiring "intelligence", such as image understanding, without >claiming to be a model of human performance of the task, then one can ask >to what extent does the work capture the underlying structure of the task? >how does the approach scale? how robust is it? and any of a number of other >questions. OK then. Point to an AI text book that covers Task Analysis? Point to work other than SOAR and ACT* where the Task Domain has been formally studied before the computer implementation? My objection to much work in AI is that there has been NO proper study of the tasks which the program attempts to simulate. Vision research generally has very good psychophysical underpinnings, and I accept that my criticisms do not apply to this area either. To supply one example, note how the research on how experts explain came AFTER the dismal failure of rule traces in expert systems to be accepted as explanation. See Alison Kidd's work on the unwarranted assumptions behind much (early?) expert systems work. One reason I did not pursue a PhD in AI was that one potential supervisor told me that I didn't have to do any empirical work before designing a system, indeed I was strongly encouraged NOT to do any empirical studies first. I couldn't believe my ears. How the hell can you model what you've never studied? Fiction. >Mr. Cockton, it is more than a little arrogant to assume that anyone who >disagrees with you is some sort of unread, unwashed social misfit When did I mention hygiene? On "unread", this a trivial charge to prove, just read through the references in AAAI and IJCAI. AI researchers are not reading what educational researchers are reading, something which I can't understand, as they are both studying the same thing. Finally, anyone who is programming a lot of the time cannot be studying people as much as someone who never programs. I never said anything about being a misfit. Modern societies are too diverse for the word to be used without qualification. Being part of a subculture, like science or academia is only a problem when it prevents comfortable interaction with people from different subcultures. Part of the subculture of AI is that the intellectual tools of maths and physics transfer to the study of humans. Part of the subculture of human disciplines is that they do not. I would be a misfit in AI, AI types could be misfits in a human discipline. I've certainly seen a lot of misanthropy and "we're building a better humanity" in recent postings. Along with last year's debate over "flawed" minds, it's clear that many posters to this group believe they can do a better job than whatever made us. But what is it exactly that an AI is going to be better than? No image of man, no superior AI. Wonder that's why some AI people have to run humanity down. It improves the chance of ELIZA being better than us. The point I have been making repeatedly is that you cannot study human intelligence without studying humans. John Anderson and his paradigm partners and Vision apart, there is a lot of AI research which has never been near a human being. Once again, what the hell can a computer program tell us about ourselves? Secondly, what can it tell us that we couldn't find out by studying people instead? -- Gilbert Cockton, Department of Computing Science, The University, Glasgow gilbert@uk.ac.glasgow.cs <europe>!ukc!glasgow!gilbert The proper object of the study of humanity is humans, not machines
smoliar@vaxa.isi.edu (Stephen Smoliar) (06/09/88)
I see that Gilbert Cockton is still judging the quality of AI by his statistical survey of bibliographies in AAAI and IJCAI proceedings. In the hope that the rest of us agree to the speciousness of such arguments, I shall try to take a more productive approach. In article <1312@crete.cs.glasgow.ac.uk> gilbert@cs.glasgow.ac.uk (Gilbert Cockton) writes: > >The point I have been making repeatedly is that you cannot study human >intelligence without studying humans. John Anderson and his paradigm >partners and Vision apart, there is a lot of AI research which has >never been near a human being. Once again, what the hell can a computer >program tell us about ourselves? Secondly, what can it tell us that we >couldn't find out by studying people instead? Let us consider a specific situation. When we study a subject like physics, there is general agreement that a good textbook must include not only an exposition of fundamental principles but also a few examples of solved problems. Why are these examples of benefit to the student? It would appear that he uses them as some sort of a model (perhaps the basis for analogical reasoning) when he starts doing assigned problems; but how doesd he know when an example is the right one to draw upon? The underlying question is this: HOW DOES KNOWLEDGE OF SUCCESSFULLY SOLVED PROBLEMS ENHANCE OUR ABILITY TO SOLVE NEW PROBLEMS? Now, the question to Mr. Cockton is: What have all those researchers who don't spend so much time with computer programs have to tell us? From what I have been able to discern, the answer is: NOT VERY MUCH. Meanwhile, there are a variety of AI projects which have begun to address the questions concerned with what constitutes experiential memory and how it might be modeled. I am not claiming they have come up with any answers yet, but I see no more reason to rail against their attempts than to attack attempts by those who would not sully their investigative efforts with such ugly artifacts as computer programs.
gilbert@cs.glasgow.ac.uk (Gilbert Cockton) (06/10/88)
In article <43@aipna.ed.ac.uk> rjc@aipna.ed.ac.uk (Richard Caley) writes: >In <1312@crete.cs.glasgow.ac.uk>, gilbert@cs.glasgow.ac.uk (Gilbert Cockton) writes >> work other than SOAR and ACT* where the Task Domain has been formally >> studied before the computer implementation? >Natural language processing. Much ( by no means all ) builds on the work >of some school of linguistics. and ignores most of the work beyond syntax :-) Stick to the computable, not the imponderable. Hmm pragmatics. I know there is AI work on pragmtics, but I don't know if a non-computational linguist working on semantics and pragmatics would call it advanced research work. >One stands on the shoulders of giants. Nobody has time to research >their subject from the ground up. But what when there is no ground? Then what? Hack first or study? Take intelligent user interfaces, hacking first well before any study of what problems real users on real tasks in real applications face (exception Interllisp-D interface, but this was an end-user project!). >According to your earlier postings, if ( strong ) AI was successful it >would tell us that we have no free will, or at least that we can not assume >we have it. I don't agree with this but it is _your_ argument and something >which a computer program could tell us. Agreed. Anything ELSE though that may be useful? (I accept that proof of our logical (worse than physical) determinism would be a revelation) >What do the theories of physics tell us that we couldn't find out by >studying objects. Nothing, but as these theories are based on the study of objects, we know that if we were to repeat the study, we would confirm the theories. Strong AI on the other hand conducts NO study of people, and thus if we studied people in an area colonised at present by hackers only, then we have no reason to believe that we would confirm the model in the hacker's mind. There is no similarity at all between the theories of physics and computational models of human behaviour, it just so happens that some (like ACT*) do have an empirical input. The problem with strong AI is that you don't have to have this input. No one would dare call something a theory in physics which was based solely on one individual's introspection constrained only by their progamming ability. In AI, it seems acceptable (Schank's work for example, can anyone give me references to the substantive critiques from within AI, I know of ones by linguists). >> The proper object of the study of humanity is humans, not machines >Well, there go all the physical sciences, botany, music, mathematics . . . And there goes your parser too. "of humanity" attaches to "the study". Your list is not such a study, it is a study of the natural world and some human artifacts (music, mathematics). These are not studies of people, OK, and they thus tell us nothing essential about ourselves, except that we can make music and maths, and that we can find structures and properties for these artifacts. A study of artifacts, cognitive, aesthetic or otherwise, is not necessarily a study of humanity. The latter will embrace all artifacts, but not as objects in themselves, but for their possible meanings. -- Gilbert Cockton, Department of Computing Science, The University, Glasgow gilbert@uk.ac.glasgow.cs <europe>!ukc!glasgow!gilbert The proper object of the study of humanity is humans, not machines
rjc@aipna.ed.ac.uk (Richard Caley) (06/13/88)
In <1342@crete.cs.glasgow.ac.uk> gilbert@cs.glasgow.ac.uk (Gilbert Cockton) >In article <43@aipna.ed.ac.uk> rjc@aipna.ed.ac.uk (Richard Caley) writes: >>Natural language processing...builds on the work of linguistics. >and ignores most of the work beyond syntax :-) Some does. >Hmm pragmatics. I know there is AI >work on pragmtics, but I don't know if a non-computational linguist >working on semantics and pragmatics would call it advanced research work. The criterion for it being interesting would not necessarily be explaining something new, explaining something in a more extensible/elegant/practical/ formal ( choose your own hobby horse ) is just as good. >But what when there is no ground? Then what? Hack first or study? Maybe my metaphor was not well chosen. Rather than building up it might be better to see the computational work building down, trying to ground its borrowed theories ( of language or whatever ) in something other than their own symbols and/or set theory. your question becomes, what when there is nothing to hang your work from? In that case you should go out and do the groundwork or, better, get someone trained in the empirical methods of that field to do it. >(exception Interllisp-D interface, but this was an end-user project!). ARGH don't even mention it, it just lost my days work for me. >(I accept that proof of our logical (worse than physical) determinism >would be a revelation) Well physical determinism wouldn't be a revelation to many of us who assume it already. I don't know your definition of logical determinism so I can't say whether that is worse. If it is meant to apply to all possible outcomes of strong AI it can't imply lack of free will ( read as the property of making your own decisions, rather than exclusion from causality ), what does it imply that is so shocking. >>What do the theories of physics tell us that we couldn't find out by >>studying objects. >Nothing. But they do. Studying objects just tells you what has happened. A (correct) theory can be predictive, can be explainatory, can allow one to deduce properties of the system under study which are not derivable from the data. >Strong AI on the other hand conducts NO study of people, Strong AI does not require the study of people, it is not "computational psycology". AI workers study people in order to avoid reinventing wheels. >>> The proper object of the study of humanity is humans, not machines >And there goes your parser too. Oops. I'm afraid I read it as parallel to "The proper study of man is man". It does seem to be something of a hollow statement; I can't think of many people who study machines as a study of humanity ( except in the degenerate case, if one believes humans are machines ). Some people use machines as tools to study human beings, some study and build machines.
jeff@aiva.ed.ac.uk (Jeff Dalton) (07/06/88)
In article <1342@crete.cs.glasgow.ac.uk> gilbert@cs.glasgow.ac.uk (Gilbert Cockton) writes:
work on pragmtics, but I don't know if a non-computational linguist
working on semantics and pragmatics would call it advanced research work.
Well, if you don't know ...