taylor@hplabsc.UUCP (Dave Taylor) (07/24/86)
This article is from "Spiros Triantafyllopoulos@386RANB" and was received on Wed Jul 23 15:10:21 1986 >>>What is the theoretical or philosophical reason that these domains could not >>>be extended to the larger but still limited ones that humans use? >> >>Nope. Expert systems have been known for behaving perfectly when faced >>with situations for which the appropriate rule/action pairs exist. However, >>expert systems can fail pretty miserably when faced with situations >>beyond the scope of their knowledge. > >"Nope" to what??? I never said that expert systems could reason outside >their domains, just that people can't either!!!! [.....] > Nope to the statement claiming the domain expansion, of course. Writing an expert system for a limited domain, such as the one DEC uses to configure VAX systems, or MYCIN, can be done; but, how about expanding said programs to more general applications within a given domain? Consider a pretty wide field, natural language understanding (where MOST of the claims came from); is there a decent expert system for understanding even a decent subset of English? The knowledge is there (i.e. syntactic and grammatic rules, meanings, etc), but not the programs for general-purpose understanding. Subset-wise, excellent programs for natural language applications (see smaller domains) such as database querying do exist. And lots of research has been going on for years. > >> People have been known for behaving (almost) perectly when faced >> with situations for which they have appropriate experience or >> knowledge. However, common sense can fail pretty miserably when >> faced with situations outside of everyday experience. >> >Is not that just as true as your statement about expert systems? Are you >really trying to say that because humans have MORE "appropriate rule/action >pairs" (i.e., knowledge or experience) that our common sense covers ALL >situations in the universe? Don't be silly! I won't bore you with >examples, because I'm sure you can come up with your own, but surely it is >obvious that there are a great many situations not covered by human common >sense. No, I'm not silly :-). Do you challenge the superior human inference capabilities, which is mostly what "common sense" is all about? In the sciences field I have found out one too many times that simple knowledge and playing by the rules is not enough to justify a certain action; inference based on that knowledge is needed. [...] >However, how is reasoning on "VERY specialized problems" different from >common sense? Our "common sense" is limited to the domain of phenomena >encountered in everyday human life, which is also very specialised in the >grand scheme of things. The domain mentioned just happens to be infinite, unless common life is reduced to living in the Iron Curtain or Orwell's society. And computers don't match with infinity. It is a very strange relationship, the one between amount of knowledge/information and display of intelligence of any sort. Expert systems perform best when confronted with HUGE amounts of information or knowledge, more than a typical human can digest at a time. Consider the VAX configurer: The last I read about it, it was keeping track of 100's of parts, configurations, requirements, specs, etc. and a rigid knowledge of how these go together. Of course, the program works fine. Compare that with a human, that has to work with much less knowledge (at a given time) available in his/her memory, (the brain may have large capacity, but, how much can it retrieve at a time?), but with superior intellect, who is then able to produce results of any worth. The expert system has more knowledge/information available to it, but much less functionality in its "inference engine". > > And since when does common sense *not* do much good >to, for example, doctors diagnosing diseases, or chemists trying to divine >the structure of organic molecules from indirect data? Again, I say that >the difference between computer "common sense" and human "common sense" is >that we haven't figured out enough of our common sense to tell it to a >computer. That's not a reflection on the computer. > For sure we don't know enough about how the brain functions or how knowledge is processed by humans. But, I will disagree about the common sense point. How about insight? when all common sense says one thing and the researcher does another, out of some kind of human feeling that is irreproducible by computers? If common sense was the prime ingredient for scientific research, we would all be Nobelaureates. Common sense is (mostly) using existing knowledge, while insight or inference is the capability of creating new knowledge. As computers are not even coming close to common sense, then geting them to infere is even furthest away. =========================================================================== Some more thoughts regarding the rabbit computer: I saw on CNN news a clip of 2 music videos urging teenagers NOT to make it. Both were for the Latin American market, where, as the commentator said, there is a big problem of teenager pregnancies. The titles were, if I am right, "when to say no" or someting and "wait". What is even more outrageous is that both were paid for by Uncle Sam Inc., thru a University (John Hopcins?) from taxpayer money. Now this is as Orwelian as it can get. I seem to remember special songs and games oriented against sexual activity in "1984".... Now we're in 1986 and we're really working our way there. So, Rabbit Co. comes late (again). Spiros Triantafyllopoulos, GM Research Labs, Warren, Mi. (Spiros@GMR.COM)
taylor@hplabsc.UUCP (Dave Taylor) (07/25/86)
This article is from pyramid!utzoo!henry (Henry Spencer) and was received on Thu Jul 24 20:33:09 1986 > ... Do you challenge the superior human inference > capabilities, which is mostly what "common sense" is all about? > > ...Of course, the program works fine. Compare > that with a human, that has to work with much less knowledge (at a given > time) available in his/her memory... but with superior intellect, > who is then able to produce results of any worth. The expert system has > more knowledge/information available to it, but much less functionality > in its "inference engine". If there is anything that the expert systems, and their development processes, have taught us, it's that inference plays a far smaller part in human thought than anyone expected. The strong point of human beings *is* the vast mass of information they have available, and the little- understood indexing system that permits rapid retrieval from a large data base by vague criteria. Well, that and the ability to add data "on the fly" based on new experience. The human brain is not a good inference engine, as witness how difficult it is to prepare a truly rigorous proof of a non-trivial theorem. Henry Spencer @ U of Toronto Zoology {allegra,ihnp4,decvax,pyramid}!utzoo!henry [Not only that, Henry, but the interesting thing about the human brain (or should I say `the mind') is that there are different classes of neural connections, or, on a more vague level, between units of information. Consider the difference between knowing that a word is spelled wrong and knowing how to spell it. When people ask me to spell a word I usually scrawl it on something near and let my pattern recog- nition system figure it out. How many times have you seen words that `looked wrong'? Just some stuff to think about...(augh! what a comment!) -- Dave]
taylor@hplabsc.UUCP (Dave Taylor) (07/27/86)
This article is from seismo!philabs!pwa-b!mmintl!franka (Frank Adams) and was received on Sat Jul 26 18:04:29 1986 >Nope to the statement claiming the domain expansion, of course. [...] >Consider a pretty wide field, natural language understanding (where MOST >of the claims came from); is there a decent expert system for >understanding even a decent subset of English? The knowledge is there >(i.e. syntactic and grammatic rules, meanings, etc), but not the programs >for general-purpose understanding. The knowledge is not there. Natural language understanding involves more than understanding the language -- one must understand the domain of discourse. Personally, I am dubious that an expert system, using anything like the current conception thereof, can be scaled up to deal with really large domains of discourse (e.g., the domain covered by common sense). But this hasn't been proven yet; no one has really tried. Frank Adams ihnp4!philabs!pwa-b!mmintl!franka Multimate International 52 Oakland Ave North E. Hartford, CT 06108