Date: Thu 4 Feb 1988 22:58-PST From: AIList Moderator Kenneth Laws Reply-To: AIList@kl.sri.com Us-Mail: SRI Int., 333 Ravenswood Ave., Menlo Park, CA 94025 Phone: (415) 859-6467 Subject: AIList V6 #26 - Connectionism, Nature of AI, Interviewing To: AIList@kl.sri.com Status: R AIList Digest Friday, 5 Feb 1988 Volume 6 : Issue 26 Today's Topics: Methodology - Two Extreme Approaches to AI & AI vs. Linguistics, Expert Systems - Interviewing Experts ---------------------------------------------------------------------- Date: 01 Feb 88 1153 PST From: John McCarthy Subject: two extreme approaches to AI 1. The logic approach (which I follow). Understand the common sense world well enough to express in a suitable logical language the facts known to a person. Also express the reasoning methods as some kind of generalized logical inference. More details are in my Daedalus paper. 2. Using nano-technology to make an instrumented person. (This approach was suggested by Drexler's book and by Eve Lewis's commentary in AILIST. It may even be what she is suggesting). Sequence the human genome. Which one? Mostly they're the same, but let the researcher sequence his own. Understand embryology well enough to let the sequenced genome develop in a computer to birth. Provide an environment adequate for human development. It doesn't have to be very good, since people who are deaf, dumb and blind still manage to develop intelligence. Now the researcher can have multiple copies of this artificial human - himself more or less. Because it is a program running in a superduper computer, he can put in science programs that find what structures correspond to facts about the world and to particular behaviors. It is as though we could observe every synaptic event. Experiments could be made that involve modifying the structure, blocking signals at various points and injecting new signals. Even with the instrumented person, there would be a huge scientific task in understanding the behavior. Perhaps it could be solved. My exposition of the "instrumented man" approach is rather schematic. Doing it as described above would take a long time, especially the part about understanding embryology. Clever people, serious about making it work, would discover shortcuts. Even so, I'll continue to bet on the logic approach. 3. Other approaches. I don't even want to imply that the above two are the main approaches. I only needed to list two to make my main point. How shall we compare these approaches? The Dreyfus's use the metaphor "AI at the crossroads again". This is wrong. AI isn't a person that can only go one way. The headline should be "A new entrant in the AI race" - to the extent that they regard connectionism as new, or "An old horse re-enters the AI race" to the extent that they regard it as a continuation of earlier work. There is no a priori reason why both approaches won't win, given enough time. Still others are viable. However, experience since the 1950s shows that AI is a difficult problem, and it is very likely that fully understanding intelligence may take of the order of a hundred years. Therefore, the winning approach is likely to be tens of years ahead of the also-rans. The Dreyfus's don't actually work in AI. Therefore, they take this "Let's you and him fight" approach by babbling about a crossroads. They don't worry about dissipating researchers' energy in writing articles about why other researchers' are on the wrong track and shouldn't be supported. Naturally there will still be rivalry for funds, and even more important, to attract the next generation of researchers. (The connectionists have reached a new level in this latter rivalry with their summer schools on connectionism). However, let this rivalry mainly take the form of advancing one's own approach rather than denouncing others. (I said "mainly" not "exclusively". Critical writing is also important, especially if it takes the form of "Here's a problem that I think gives your approach difficulty for the following reasons. How do you propose to solve it?" I hope my forthcoming BBS commentary on Smolensky's "The Proper Treatment of Connectionism" will be taken in this spirit.) The trouble is "AI at the Crossroads" suggests that partisans of each approach should try to grab all the money by zapping all rivals. Just remember that in the Communist Manifesto, Marx and Engels mentioned another possible outcome to a class struggle than the one they advocated - "the common ruin of the contending classes". ------------------------------ Date: 3 Feb 88 02:37:00 GMT From: alan!tutiya@labrea.stanford.edu (Syun Tutiya) Subject: Re: words order in English and Japanese In article <3725@bcsaic.UUCP> rwojcik@bcsaic.UUCP (Rick Wojcik) writes: >Still, >it seems to have little to do with the problems that AI researchers busy >themselves with. And it has everything to do with what language >scholars busy themselves with. Perhaps the participants realize >instinctively that their views make more sense in this newsgroup. I am no AI researcher or language scholar, so find it interesting to learn that even in AI there could be an argument/discussion as to whether this is a proper subject or that is not. Does what AI researchers are busy with define the proper domain of AI research? People who answer yes to this question can be safely said to live in an established discipline called AI. But if AI research is to be something which aims at a theory about intelligence, whether human or machine, I would say interests in AI and those in philosophy is almost coextensive. I do not mind anyone taking the above as a joke but the following seems to be really a problem for both AI researchers and language scholars. A myth has it that variation in language is a matter of what is called parameter setting, with the same inborn universal linguistic faculty only modified with respect to a preset range of parameters. That linguistic faculty is relatively independent of other human faculties, basically. But on the other hand, AI research seems to be based on the assumption that all the kinds of intellectual faculty are realilzed in essentially the same manner. So it is not unnatural for an AI researcher try to come up with a "theory" which should "explain" the way one of the human faculties is like, which endeavor sounds very odd and unnatural to well-educated language scholars. Nakashima's original theory may have no grain of truth, I agree, but the following exchange of opinions revealed, at least to me, that AI researchers on the netland have lost the real challenging spirit their precursors shared when they embarked on the project of AI. Sorry for unproductive, onlooker-like comments. Syun (tutiya@csli.stanford.edu) [The fact that I share the nationality and affiliation with Nakashima has nothing to do with the above comments.] ------------------------------ Date: 30 Jan 88 17:34:31 GMT From: trwrb!aero!venera.isi.edu!smoliar@ucbvax.Berkeley.EDU (Stephen Smoliar) Subject: interviewing experts I just read the article, "How to Talk to an Expert," by Steven E. Evanson in the February 1988 issue of AI EXPERT. While I do not expect profound technical insights from this magazine, I found certain portions of this article sufficiently contrary to my own experiences that I decided a bit of flaming was in order. Mr. Evanson is credited as being "a practicing psychologist in Monterey, Calif., who works in the expert systems area." Let me being with the observation that I am NOT a practicing psychologist, nor is my training in psychology. What I write will be based primarily on the four years of experience I had at the Schlumberger-Doll Research Laboratory in Ridgefield, Connecticut during which I had considerable opportunity to interact with a wide variety of field experts and to attempt to implement the results of those interactions in the form of software. Mr. Evanson dwells on many approaches to getting an expert to explain himself. For the most part, he address himself to the appropriate sorts of probing questions the interviewer should ask. Unfortuntely, one may conclude from Mr. Evanson's text that such interviewing is a unilateral process. The knowledge engineer "prompts" the expert and records what he has to say. Such a practice misses out on the fact that experts are capable of listening, too. If a knowledge engineer is discussing how an expert is solving a particular problem; then it is not only valuable, but probably also important, that the interviewer be able to "play back" the expert's solution without blindly mimicking it. In other words, if the interviewer can explain the solution back to the expert in a way the expert finds acceptable, then both parties can agree that the information has been transferred. This seems to be the most effective way to deal with one of Mr. Evanson's more important observations: It is very important for the interviewer to understand how the expert thinks about the problem and not assume or project his or her favored modes of thinking into the expert's verbal reports. Maintaining bilateral communication is paramount in any encounter with an expert. Mr. Evanson makes the following observation: Shallowness of breathing or eyes that appear to defocus and glaze over may also be associated with internal visual images. Unfortunately, it may also indicate that the expert is at a loss at the stage of the interview. It may be that he has encountered an intractable problem, but another possibility it that he really has not processed a question from the interviewer and can't figure out how to reply. If the interviewer cannot distinguish "deep thought" from "being at a loss," he is likely to get rather confused with his data. Mr. Evanson would have done better to cultivate an appreciation of this point. It is also important to recognize that much of what Mr. Evanson has to say is opinion which is not necessarily shared "across the board." For example: As experts report how they are solving a problem, they translate internal experiences into language. Thus language becomes a tool for representing the experiences of the expert. While this seems rather apparent at face value, we should bear in mind that it is not necessarily consistent with some of the approaches to reasoning which have been advanced by researchers such as Marvin Minsky in his work on memory models. The fact is that often language can be a rather poor medium for accounting for one's behavior. This is why I believe that it is important that a knowledge engineer should raise himself to the level of novice in the problem domain being investigated before he even begins to think about what his expert system is going to look like. It is more important for him to internalize problem solving experiences than to simply document them. In light of these observations, the sample interview Mr. Evanson provides does not serve as a particularly shining example. He claims that he began an interview with a family practice physician with the following question: Can you please describe how you go about making decisions with a common complaint you might see frequently in your practice? This immediately gets things off on the wrong foot. One should begin with specific problem solving experiences. The most successful reported interviews with physicians have always begun with a specific case study. If the interviewer does not know how to formulate such a case study, then he is not ready to interview yet. Indeed, Mr. Evanson essentially documents that he began with the wrong question without explicitly realizing it: This question elicited several minutes of interesting unstructured examples of general medical principles, data-gathering techniques, and the importance of a thorough examination but remained essentially unanswered. The question was repeated three or four times with slightly different phrasing with little result. >From this point on, the level of credibility of Mr. Evanson's account goes downhill. Ultimately, the reader of this article is left with a potentially damaging false impression of what interviewing an expert entails. One important point I observed at Schlumberger is that initial interviews often tend to be highly frustrating and not necessarily that fruitful. They are necessary because of the anthropological necessity of establishing a shared vocabulary. However, once that vocabulary has been set, the burden is on the knowledge engineer to demonstrate the ability to use it. Thus, the important thing is to be able to internalize some initial problem solving experience enough so that it can be implemented. At this point, the expert is in a position to do something he is very good at: criticize the performance of an inferior. Experts are much better at picking apart the inadequacies of a problem which is claiming to solve problems than at giving the underlying principles of solution. Thus, the best way to get information out of an expert is often to give him some novice software to criticize. Perhaps Mr. Evanson has never built any such software for himself, in which case this aspect of interacting with an expert may never have occurred to him. ------------------------------ End of AIList Digest ********************