The MacNeil/Lehrer Report; Artificial Intelligence

[Tease]

PENNY NII, computer scientist: It's intelligent because it has knowledge, and it's also knowledge not of just textbook knowledge, but knowledge based on experience. And it does have a line of reasoning that it follows, and it can explain itself. And so all those behaviors I would think that if a person did it, you would say that was intelligent. That's exactly what the computer is doing.

[Titles]

ROBERT MacNEIL: Good evening. In the science fiction movie 2001 the space mission was run by a supremely intelligent yet soft-spoken computer named Hal. Well, it's only 1983, but the era of the intelligent computer is already upon us. At a handful of universities and high-tech companies scientists are creating a new generation of computer programs that can mimic human thought. They are capable of making judgments, giving advice, even learning from their own mistakes. The field is called artificial intelligence, and its potential has excited corporations and many professions. Visionaries see such computers doing a lot of what highly skilled humans now do -- doctors, lawyers, stockbrokers, tax consultants, air traffic controllers and so on. The Europeans and Japanese are pushing ahead in the field. The boldest effort is in Japan, which has committed itself to deliver by 1990 a computer that will converse easily with people in ordinary language. But the dawn of intelligent computers raises many questions. If we entrust important decisions to a machine, could it go haywire, like Hal in 2001, and harm us? Tonight, artificial intelligence: the potential and the limits. Jim?

JIM LEHRER: Robin, the traditional everyday use of computers is as a repository of facts and figures, each available for immediate retrieval at the push of a few buttons. What is former Secretary of State Vance's first name? "Cyrus," answers the computer, because it has been programmed with that fact. But the artificial intelligence computer goes beyond that -- dramatically beyond. It is programmed with a general set of guidelines or principles so it can make sophisticated judgments. Cyrus Vance was in fact the subject of such an experiment at Yale University. A computer was programmed with a multitude of news stories about Vance. Then it was asked the question, "Did Cyrus Vance's wife ever meet Menachem Begin's wife?" Nowhere in its clipping file in its memory bank was there a specific statement that the two women had met, but the computer answered, "Probably yes." Why? Well, because the computer knew Vance and Begin had attended state dinners together and it knew the principle that wives often accompany their husbands to such affairs. So, like an ordinary mortal, the computer put two and two together and reasoned. One of the most practical applications of this is in the so-called expert computer where the expertise of a recognized authority in geology, medicine or whatever is programmed in, making the computer able to reason much like the expert himself would do. Our producer Ken Witty and reporter Maura Lerner went recently to northern California to see such an experiment in the medical field.

MacNEIL: [voice-over]: James Kennedy is a lung patient at Pacific Medical Center in San Francisco. Here at the hospital's pulmonary labs he is undergoing a series of tests that gauge whether he might have an obstructive disease such as emphysema.

DOCTOR [voice-over]: Deep breath, deep breath, deep breath. Blow it out, blow it out, blow it out. Use your stomach now; squeeze out. Squeeze it out a little bit more.Keep pushing, keep pushing, keep pushing. One big last push.

MacNEIL [voice-over]: After Mr. Kennedy breathes in and out on the machine, a computer takes over. First it calculates his lung performance; then another computer program interprets the test results, just as a doctor would. The computer program, called PUFF, issues a printout diagnosing Mr. Kennedy's condition. His tests and case history indicate emphysema. PUFF's diagnostic abilities are patterned after the knowledge and experience of Dr. Robert Fallat, the hospital's chief of pulmonary medicine.

Dr. ROBERT FALLAT: By and large PUFF is acceptable, accurate, usable the way it comes out the first time in 75 to 85 percent of the cases. Actually, you know, 85% is really nice to be able to have a finished product, and 85% of the times you're willing to let it to out as a final report.

MacNEIL [voice-over]: PUFF is the product of a new kind of collaboration between doctors like Bob Fallat and computer scientists like Penny Nii.The idea is to design computer programs that capture the knowledge and reasoning processes of recognized authorities. Professor Nii spent three months with Dr. Fallat learning how he worked and thought. In his case, understanding how he approached his work wasn't easy at the beginning.

PENNY NII: He wasn't doing it the way he described it, and he wasn't doing it the way that he thought he was doing. I think he was very surprised that it was different than he thought because we have seen this with other experts, too, that when they explain what they're doing to other people it's perhaps ideal situation or textbook explanation which any physician is used to. But it's very hard to know what you've learned from experience.

Dr. FALLAT: Penny Nii kept -- and all the group down there at Stanford kept trying to pin me down as to exactly what I look at next, you know. What do I look at first? What do I look at next? What do I really put together to make a conclusion? And I really had to rethink my own thinking, you know, to understand, you know, what am I really doing, you know. What sequence do I really follow?

Prof. NII: Expert can go on and on and on forever explaining the general theories, general knowledge, but what we really want are specific rules for specific situations.

MacNEIL [voice-over]: Penny Nii discovered that Dr. Fallat used some 30 rules based on his clinical experience to diagnose whether patients have obstructive airway disease. Once the basic medical rules were established, it took two student programmers a matter of weeks to come up with the first crude model of PUFF.

Dr. FALLAT: There's a lot of what we do, including our thinking and our expertise, which is routine, and which doesn't require any special human effort to do. And that kind of stuff should be taken over by computers. And to the extent that 75% of what I do is routine and which all of us would agree on, why not let the computer do it and then I can have fun working with the other 25%.

MacNEIL [voice-over]: If PUFF can diagnose Dr. Fallat's lung patients as well as he can three-quarters of the time, is the machine an example of artificial intelligence?

Prof. NII: Yes, of course it's intelligent because it has knowledge, and it's also knowledge not of just textbook knowledge, but knowledge based on experience. And it does have a line of reasoning that it follows, and it can explain itself. And so all those behaviors, I would think that if a person did it, you would say that was intelligent. That's exactly what the computer is doing.

MacNEIL [voice-over]: PUFF is a spinoff from years of research at Stanford University and other centers of artificial intelligence. The operational center for this work is this bank of computers on the Stanford campus. Using this hardware, Dr. Edward Shortliffe and a team of computer scientists spent several years in the mid-1970s figuring out ways for a computer program to pursue a line of reasoning in the same way a medical specialist does. The research program which became the model of their thinking is called MYCIN.It was created to diagnose blood infections and meningitis cases and then help doctors choose the right drug for treatment. It operates on a knowledge base of 500 rules gathered from experts in infectious disease.

Dr. EDWARD SHORTLIFFE, Stanford University: The program actually guides the interaction and I've started a little case here of the make-believe patient by the name of James Johnson, a brand-new patient who has just come into our emergency room with a stiff neck and a fever, and we are suspicious of meningitis. Now the program, with that basic information, if beginning to consider the possibility that in fact this patient does have meningitis. Therefore it wants to know something about signs and symptoms of that disease. It asks here about symptoms, complaints that the patient has, such as dizziness, lethargy, stiff neck, headache. And in fact this patient does have a persistent headache, so from this point on questions will focus on the possibility of a meningitis case.

MacNEIL [voice-over]: The program's expertise has already won impressive recognition from medical peers. When MYCIN was pitted against five human specialists in infectious disease, the computer won highest marks for its treatment of meningitis.A special feature built into MYCIN is the ability for doctors to ask the computer to explain its conclusions.

Dr. SHORTLIFFE: Every question, since it has some rule leading to it, can respond to the demand from the user, "Why?" Whatever rule is currently under consideration can be translated into English and gives a reasonable explanation to the physician of why that piece of information is useful. From the user's point of view what makes this artificial intelligence is the ways in which MYCIN's performance mimics or simulates what one would expect from a human consultant in this field. One gets the feeling after seeing MYCIN run on a few cases that there is a level at which it understands what's going on in this field, that it has some knowledge that goes beyond some kind of rote description of a simple formula one must always use in treating of patients. It seems to be able to adapt to subtleties in the way that we expect an expert to be able to do.

MacNEIL [voice-over]: Despite its obvious talents, MYCIN turned out not to be practical. It demanded too much time from the doctors. If they wanted time on a centralized computer they had to wait their turn. For doctors in a hurry that was frustrating. One solution is a new generation of mini-computers. These small but powerful machines will soon make it possible for doctors and other users to have their own computers. The new machines, with flashy graphics, are much easier to use. You don't even have to know how to type. Another lesson learned from the MYCIN research at Stanford is that for a computer to be used in medicine it has to become part of the daily activities of the staff. That's what happens here at the cancer ward at Stanford Medical Center. Here, each patient receives an individualized series of treatments involving radiation therapies and experimental drugs.

Dr. ROBERT CARLSON: Do you have any trouble with nausea and vomiting with the chemotherapy?

PATIENT: Nausea, a little, but no vomiting.

Dr. CARLSON: Good.

MacNEIL [voice-over]: After Dr. Robert Carlson consults with the patient, he goes to the computer terminal to update her chart. He enters the information and the program known as ONCOCYN suggests what dosages of various drugs the patient should receive next. Dr. Charlotte Jacobs is head of the clinic, and worked with Dr. Shortliffe in the development of ONCOCYN:

Dr. CHARLOTTE JACOBS: We see large volumes of patients here, about 700 patientvisits a month. Many of our patients are on what are called treatment protocols, or research protocols, which means that the treatments are not only complex but that the patients see various tests done at particular periods in time. I initially thought that if I had a patient come in and I wanted a particular treatment, I would punch it into the system and it would tell me what the treatment was. So that it essentially would do what a book would do.Instead of flipping through a book, instead of what it really has ended up doing, which is keeping a running record of the patient and advising, making complicated calculations on patient treatments, educating, so that the system has really done a lot more than I initially thought possible.

MacNEIL [voice-over]: But how do the doctors who work with them on a daily basis react to having computers give them advice?

Dr. CARLSON: It certainly is useful for storing data.As far as making suggestions, I think it's early and it's a simple system. As the system becomes more sophisticated I think it will have a greater effect in decision-making.

DOCTOR: We will never rely on the computer to calculate the exact dose that we're going to give the patient, but it's a good doublecheck against what we decide that we're going to give the patient.

Dr. FALLAT: There's also some resistance, I guess, just like there always is in anything new, that the idea that the computer is going to be taking over your job. And are you still going to be able to charge for interpreting tests if the computer is doing it for you? In the end, the use of computers and the use of artificial intelligence or any kind of computer is in fact going to be a way to, in the end, really improve our practice of medicine because the computers will be able to take this mass of technological information, crunch out the nembers, look at the numbers, tell us what's important, what isn't important, do the drudgery parts of it, and allow us to spend more time doing the human parts of medicine.

MacNEIL: Now to a man who's an expert on these expert systems. He is Randall Davis, assistant professor in the artificial intelligence lab at MIT. Professor Davis, just how good are these medical diagnosis systems today?

RANDALL DAVIS: Well, I think it's important to understand that the question is not so much, how good are the systems?The question really is, how good is our understanding of the tasks and the problems that we're trying to solve? One example we've seen just a minute ago is this pulmonary function problem, and apparently our understanding of how to do diagnosis of pulmonary function is good enough that it can be written down in a selection of some 30-odd rules.There are other problems, like the MYCIN system, where it takes 500-odd rules. But what we have in AI is a technology, a way of capturing judgmental rules in a machine --

MacNEIL: In other words, if a doctor for instance is good enough at figuring out the way he does it in a systematic way, the computer capacity is up to making a diagnosis?

Prof. DAVIS: Yes, indeed. The real focus is on understanding logic, on understanding inference and on understanding how the problem-solving is being done. We have a certain technology available for doing that, and the question is, how much do we understand about the world around us? What of the problems that we routinely encounter can be written down in this fashion?

MacNEIL: What can't computer programs do at the moment?

Prof. DAVIS: Well, one of the classic problems that stands in our way of building lots of these systems and applying them widely is learning -- the knowledge acquisition problem is the classic bottleneck here. PUFF is an exception in the sense that it's a reasonably small problem. It was built over a matter of several man-months of effort. MYCIN is a good example, with those 500 rules extending over, oh, 10 to 15 man-years of effort.

MacNEIL: Am I putting it too simply if I ask you that if the computer is going to learn something, the person who is programming it has got to tell it that additional thing?

Prof. DAVIS: Currently that's the state of the art, indeed.

MacNEIL: All right. Is it going to move to the point where the computer can learn spontaneously from putting things together?

Prof. DAVIS: We're working on it. One of the active fields of research is to see if we can develop the principles of learning from experience. People clearly do it. Doctors indeed are taught much in that fashion. Medical education is looking at and dealing with lots and lots of patients and, in effect, deriving the rules by which one ought to do it, or deriving experience by which one ought to do it, by learning from those cases. We would like to be able to do that.

MacNEIL: Is one of these medical diagnosis expert systems now as good as an inexperienced doctor? For instance, in a small rural practice, where maybe a doctor is fresh out of medical school without much experience, would the patients there be better off with the computer than with the doctor? Or the doctor plus the computer?

Prof. DAVIS: The doctor plus the computer, perhaps. I think it's reasonable to say that any time we can supply people with tools that will help them, then indeed that's an excellent thing to do. Whether it's better than a particular individual -- whether one particular program is better than a particular individual is hard to answer in general, of course. But to the extent that we can supply decision assistance tools that can, as were indicated in the film, be used as a second opinion, in effect, that can be a very effective mechanism.

MacNEIL: How soon is it going to be before these expert systems or artificial intellignece programs really begin to change our lives, the way we do things?

Prof. DAVIS: I don't think we're going to see a watershed day. It isn't going to be announced; there isn't going to be a sudden threshold, and it won't be in the newspapers. I think we're seeing, as in all such things, evolution. The existence of personal computers and some of the things we might begin to do with them beyond simple bookkeeping -- it's going to happen slowly, and in very small ways it's beginning to happen. The simplest example that people are probably familiar with at the moment is the existence of chess-playing programs. I mean, that would have been unheard of 25 years ago. That was a research effort 25 years ago, and yet now it's simple consumer product that isn't all that expensive. What we're going to see is a whole lot of very small increments like that, that in hindsight will add up to quite a change.

MacNEIL: Thank you. Jim?

LEHRER: No discussion about artificial intelligence or a story about it goes very far before the name Roger Schank crops up. He's director of the artificial intelligence laboratory at Yale University, and he also heads a company which is in the business of marketing artificial intelligence concepts. Dr. Schank, first of all, what do you think of expert system that we've just seen a tape on and Professor Davis has just explained?

ROGER SCHANK: Well, I think they're very nice, but I have my doubts about their ultimate significance.

LEHRER: Why? What are your doubts?

Prof. SCHANK: The real hard problem, the issue of whether machines are intelligent -- the real hard problem is that intelligence is the putting together of different diverse experiences to come up with new ideas. And what you have in the knowledge engineering efforts is an attempt to codify knowledge as it currently exists but not be able to relate a to b. So, for example, the expert in meningitis wouldn't necessarily be able to relate that knowledge to some other area, whereas a doctor might be able to do that. And what we really need in artificial intelligence are rules for abstraction and learning and reasoning from particular experiences, and you don't see any of that in the expert systems work.

LEHRER: All right, now, you're involved in another area. For instance, I mentioned the Cyrus Vance thing at Yale. That was your project. You also did one trying to duplicate the mind of a football coach. What's your thing all about, and what purpose do those kinds of things serve?

Prof. SCHANK: Those have two different purposes. The football coach is an attempt -- you have to understand, artificial intelligence is essentially a scientific enterprise in the university laboratory. It has its commercial aspects, which I can talk about, but from the point of view of the football coach what we're looking for is a computer program that will model reasoning on the fly from experience that'll enable you to get better. The premise is that a football coach calls a play, all of a sudden, that he says is pulled out of intuition. He thought it would work. And we want to get at, how does he create that intuition? How does he come up with a new idea? How does he reason from a past experience and remember that 15 years ago in the Harvard game this thing worked, and "let me try it again"? That's the kind of reasoning we're looking for.

LEHRER: Well, is the ultimate goal to replace the football coach or to help the football coach?

Prof. SCHANK: Well, actually I'm not very interested in either of those. I'm much more interested in finding out how the football coach or any expert reasons from his experiences. I think it'd be rather fun to have a computer calling the plays in a game, but I'm not sure how important that is.

LEHRER: Well, what is the practical application, as you see it, down the line for what you're doing and what everybody else is doing in the artificial intelligence area?

Prof. SCHANK: There are a number of practical applications with respect to the Cyrus program that you mentioned. One of the problems that we have in the computer business is there are lots of data bases around that are basically very dumb. It's very difficult to get at them. You have to ask the question in a specified way and if youdon't ask it that way it won't understand, and it could have a piece of information as simple as "Shakespeare wrote Hamlet," and you can be foolish enough to say, "Who was the author of Hamlet?" and have it not understand that author and writer are related. And in the Cyrus example, just being able to reason from two different pieces of information and pull in a new one. So what we're looking for is --

LEHRER: Yes, but what's the value of that?

Prof. SCHANK: The value of that is that at some point you can walk up to a computer and ask for advice and get some.

LEHRER: Well, but why is that important? Why couldn't you just go -- why is that better than going and asking a human being the same question?

Prof. SCHANK: It isn't. The only difference is that at a branch office of a bank in Iowa you may not have the very best financial expert available to a man who makes $30,000 a year, but we can take that vested financial expert, model him on a machine, give it a natural-language capability, and allow somebody to come in, type his question and get some good advice.

LEHRER: You do see commercial applications of this.

Prof. SCHANK: I think that's one of them.

LEHRER: And your company is moving in that area?

Prof. SCHANK: My company is interested in building natural-language programs that provide advisory services to the general public that otherwise wouldn't be available; for example, financial advice.

LEHRER: "Natural language" meaning the kind we're talking right now, rather than computer language, right?

Prof. SCHANK: That's right. Meaning you can sit there and type anything you want and it'll understand what you asked.

LEHRER: Thank you. Robin?

MacNEIL: Dr. Davis, what about Professor Schank's doubts about the ultimate significance of the expert programs like the medical diagnosis one?

Prof. DAVIS: Well, I think that gets down to a technical disagreement that I think Roger's construing a little bit too narrowly what he's just seen. Indeed, that technology as it exists is the thing which is available for use in the hospitals these days, but that's not all there is to the technology, and indeed, the problems that he talks about -- about learning from experience, about knowing -- he said something very interesting. He said, what are the rules of abstraction? What, in fact, is the skill behind knowing how to learn? And indeed that's one of the active areas --

MacNEIL: You doubt that that can be -- that that can be programmed?

Prof. SCHANK: No, I'm working -- my whole laboratory is devoted to trying to solve that problem.

MacNEIL: Well, I don't understand the basis of your doubt.

Prof. SCHANK: Well, I think the only doubt I have is that the expert systems as they're currently constituted are much more limited than, say, the public or the press might have led people to believe, and that there's major scientific breakthroughs are still needed before we really have genuine experts available by computer.

MacNEIL: What is going to be the really important application of artificial intelligence? Is it going to be in providing consultancy for specialists like doctors and others, or is it going to be in providing advice for laymen?

Prof. SCHANK: I think the issue is that the man in the street doesn't get significant information. He gets a lot of things thrown at him, but when he wants to know whether he should sell his house or he wants to know whether he should invest in a tax shelter or he wants to know any kind of information -- should he buy an insurance policy, he doesn't have anybody that he really can ask who has the best possible advice who can give it to him.

Prof. DAVIS: I think that there are high-leverage problems and there are the problems of less economic significance, and I think what we're going to see is, again, an evolution. Right now, because the technology is still reasonably expensive and difficult to put together, the problems that people are working on are those with the highest economic payoff, which typically tend to be very narrowly defined kinds of expertise in specialized industries, like the oil industry, for example. To the extent that the technology becomes easier to build and to deal with, then in fact we'll begin to see it evolve out into less specialized issues of the sort Roger is referring to.

MacNEIL: Is what you mean that there is a lot of expertise in this country, but at the very top level it's expensive and rare, and then there is a lot of expertise of less quality which is less expensive and more accessible? Is it possible for this artificial intelligence kind of programming to bring that very top expertise down so that lots of people can have access to it?

Prof. SCHANK: That's precisely its value. Not only -- if we could build a model of Henry Kaufman, you could have a conversation with him and test his --

MacNEIL: Well, I can have a conversation with him. We have.

Prof. DAVIS: Note also that it doesn't matter that we ever achieve absolute top expert performance or indeed superhuman performance, as some of the critics of the field have claimed that we'll never get to being better than people. That doesn't matter. There's an enormous amount of intellectual payoff; there's an enormous amount of utility. There's an enormous amount of good to be gained from having even moderate levels of expertise available.

MacNEIL: But what about bringing the expertise of a top medical specialist, say, at Johns Hopkins, available to an intern working late at night in a hospital faced with a difficult diagnosis? Can it do something like that ultimately?

Prof. DAVIS: That's our aspiration.Ultimately we'd like to do exactly that.

MacNEIL: Are you doubtful that it can do that?

Prof. SCHANK: No, I think that's possible. To some extent it can capture, say, 75% of his expertise, but it's that 25% that it can't capture that may be the most important part of his expertise.

MacNEIL: Of the specialist? The top person.

Prof. SCHANK: That's right. His ability to recognize a situation that no one else could see, not because he had some rule, but because he just had a feeling about it. See, these feelings that we talk about as human beings, there are really some cognitive processing going on.

MacNEIL: I get the impression from both of you that we are a long way from Hal running a space mission.

Prof. DAVIS: Agreed.

MacNEIL: Really a long way. I mean turn of the century or beyond that before it can really mimic human thought processes.

Prof. DAVIS: I think there's an important kind of calibration we have to do here, which is to recognize that the interest and excitement and the sense of the pace of change in this field is, in large measure, a reflection of a migration of a set of research ideas out into the general public. The changing economics of the hardware makes it possible to take a number of research ideas that have been around awhile and move them out into the larger world to get to those truly speculative things. It's going to take much more significant intellectual break-throughs, and those proceed at a much slower pace.

MacNEIL: Well, thank you, Professor Davis, Professor Schank. Good night, Jim.

LEHRER: Good night, Robin.

MacNEIL: That's all for tonight. We will be back on Monday night. I'm Robert MacNeil. Good night.