Probably the most famous early idea about equating machine intelligence and the human mind is a notion called the Turing test. It derives from a famous paper by the I guess you could call him the mathematician and computer scientist, Alan Turing. In 1950, he wrote a paper called Computing Machinery and Intelligence, and it was published in the journal Mind, philosophical journal. In that paper, he outlined the idea of an operational test for whether you could say that a computer or machine was intelligent or had a human-like intelligence. Alan Turing along with John von Neumann, with whom his name is usually paired. Alan Turing is one of the two great founding figures of computer science. In 1936, '37 he wrote a paper, which introduced the idea of a Turing machine, which was a mathematical abstraction supposed to represent the idea of a computer. Well, at that point, what people generally meant by a computer was a human being, a person whose job it was to actually do a certain routinized arithmetic in order to compute tables and so forth. There were no computers of the sort that we think of as existing at that point. That is to say there were no automatic computers, there were no electronic computers. But when Alan Turing wrote his paper, he was providing a mathematical abstraction of any computing machinery, and his idea in that paper was to try and formalize the idea of what it meant to be computable by mechanical means, you could put it that way. There was a similar motivating force behind his later 1950 paper, which introduced the idea of the Turing test. He didn't call it the Turing test in his 1950 paper. It's been called that since that time but in his paper, he called it The Imitation Game. The basic idea is this; that you put a computer and a person behind two separate doors in two separate rooms, and the doors are closed and you as a judge come up to these two doors and you can type questions at the time of the writing of his paper in 1950, it would have been something like a teletype. You could imagine typing over a computer line to both rooms. You can get back textual answers from both rooms. One is a person just answering questions the way a person would. They have no particular instructions except to answer the questions however they want. The other room contains a computer whose job it is to fool you the judge into thinking that it is a person. Your job is to tell the two apart. If you can't tell the two apart, if you can't say, so here's room number one with the person and here's room number two with the computer, what shall I do? You would not see this as far as you're concerned as the judge, you just see two doors, but behind one of these doors is a person and behind the other is a computer, which I'll just represent as a screen and the keyboard. So your job is to tell these two things apart via the answers that they provide over the computer connection or teletype or whenever. If you can't do that in about 15 minutes or so, then as far as Turing is concerned in this paper, you could say of the computer that it is intelligent. It certainly has a kind of operational intelligence good enough to fool a person into thinking that this is a human. Now, when he introduced this idea in his paper, he actually begins with a scenario that is similar but not identical to this. He imagines an imitation game played between a man and a woman. So behind one door is a man, behind the other door is a woman. You can type questions at both of them. You don't know which room contains which person. So you can type questions to both of them and your job. Well, without loss of generality, you could say that the man's job is to pretend to be a woman and the woman's job is just to answer the question however she wishes. Your job as judge is to tell the two rooms apart. Say, which of the rooms contains a woman answering the questions and which of the rooms contains a man pretending to be a woman. Notice that in this scenario, which by the way, as far as I know, I don't know if anyone has actually done this experiment. Sounds like it could be a fun party game or something like that. But regardless, there are two challenges that has to be met here. One is the man whose job is to successfully pretend to be a woman, the other challenge is that of the judge whose job is to tell the two apart. You could imagine doing versions of this imitation game tests with a lot of other pairings. You could have say, a 20-year-old and a 70-year-old, where the 20-year-old's job is to pretend to be the older person and your job as judge is to tell which of the two rooms contains a genuine 70-year-old person and which contains a young person pretending to be an older person or you could do this with different geographic locations. Somebody from Alaska and somebody from South Carolina or you could do it with different political orientations. There are a whole variety of things. In any of those cases, it might make for a really interesting experiment. In the test that Turing imagines, the job is to tell a person from a computer. Early on in the paper, he provides a sample question and answer that the judge might give, where the judge is asking questions and he or she is getting back answers from one of the two rooms, and the judge's job is to determine whether this might be a computer or a person. So I'll just read it out. Please write me a sonnet on the subject of the fourth bridge. Answer: Count me out on this one. I never could write poetry. Add couple of numbers, wait for awhile and then response is printed out. Do you play chess? Yes, I have king at my K1 and no other pieces. You have only a king at K6 and rook at R1, that's your move, what do you play and so forth. Even in this little exchange, there's something rather clever and charming going on, which is, if you examined the addition problem carefully, you will see that whoever is answering this question, whether it's a person or a computer, whoever is answering this question, gives the wrong answer. Now, you might reason if you were the judge that superficially, you might think, "Well, that has to be the person, not the computer answering the question because after all, a computer would not get the answer to an addition problem wrong, would it?" Well, remember this is a computer programmed to imitate a human being. Therefore, it might be a computer whose program tells it to make the errors that people are likely to make in addition. In fact that the error that's made here is not a wildly off base error, it might be the kind of error that a person would actually make in adding two numbers. So this is a very cute introduction to the paper. The idea again is that Turing is not saying that if a computer can pass this test, it's a human, nothing like that. All that he's arguing is that as far as we're concerned, we could think of this test as a way of just in a concrete way settling the argument about whether a machine could actually be intelligent. If a machine could pass this test, then it would be Intelligent. In the course of the paper, Turing provides a bunch of potential objections to his idea. You get the feeling that in writing this paper in 1950, Alan Turing went around to a variety of people, and talked to them about this idea and got back certain objections and answers, and he wrote these all out and provided his own responses to these objections. As I give this lecture, I'm hoping that you read the paper or have read it before. You're actually watching this lecture, and you've already made some internal decisions for yourself about the provocative myths or interest of these various objections. I'm not going to go through all of them just for lack of time, but I'll mention several. People have different favorites or they want to focus on different objections. I'll mention a few of them. The argument from consciousness, Lady Lovelace's objection and the argument from ESP, all interesting. In fact, all the objections could be things that we could talk more about but I'll just focus on these three. So what about the argument from consciousness? A lot of people would look at Turing's test and say a bit, sure, you could say that maybe a computer could pass this test but you wouldn't say that the computer was conscious or had anything like human consciousness in doing this. Turing's response to this in a way that may or may not convince you, he says something to the effect that, "Well, consciousness is quite mysterious. We don't really know whether it involves. But again, if you were having a conversation with the computer and it looked something like the following," and he gives this example. This is the question. In the first line of your sonnet which reads, "Shall I compared thee two a summer's day?" Would not a spring day do as well or better? Answer, it wouldn't scan. How about a winter's day? That would scan all right, yes. But nobody wants to be compared to a winter's day. Would you say Mr. Pickwick reminded you of Christmas? Answer, in a way. Yeah, Christmas is a winter's day and I do not think Mr. Pickwick would mind that the comparison. I don't think you're serious. By a winter's day, one means the typical winter's day rather than a special one like Christmas. Now, look, if you were having that conversation with an entity behind the door, you would be willing to attribute, if not consciousness, then certainly a great deal of knowledge to this entity. It's read Dickens, it knows about Mr. Pickwick and it knows that Christmas is a special winter's day, and it knows about the rhythm of lines and poetry. It feels like a very natural conversation. In fact, bordering on an interesting conversation, the one I was having with this entity. So Turing's argument is well, if you could have a conversation of this kind, then for whatever it's worth, you'd probably be willing to attribute intelligence and perhaps consciousness to. It might be very difficult to tease apart whether the entity really did have consciousness or not. But certainly, in an offhand and formal way, you would be willing to attribute consciousness to it. If I say this may or may not convince you as an argument, but that's the response that Turing gives. In the case of Lady Lovelace's objection, that's an interesting objection. He quotes a paper that was written in the early 1840s by Countess Ada Lovelace. Countess Ada was a colleague, a collaborator of the British mathematician, Charles Babbage. Charles Babbage, this a fascinating story and I only wish I had more time to go through the entire history of it because it really is an amazing story. Charles Babbage was an early 19th century genius, a mathematician. He had many interests. Over his life, his overriding desire, his obsession became to build something which he called the analytical engine. Analytical engine being a mechanical device that could automatically do a wide, it in fact has a strong resemblance to a general purpose computer. If you read about Babbage's life and you read some of his notes, it's very clear that he had many of the pioneering ideas about how a computer might work, and what a computer could do although in his vision, he was thinking of this as a mechanical, perhaps steam-driven entity rather than of naturally this being in the early 18 hundreds rather than being an electrical device. Nonetheless, his ideas about computing were quite amazing. He never did actually get to build the analytical engine because the engineering of the time was not up to the cost and precision that would have been required to build a working analytical engine. Although, there's more to say about that as well. In any event, Babbage worked with Countess Ada Lovelace, who was the daughter of the famous poet, Lord Byron. She actually co-wrote along with a very cogent Italian writer named Luigi Menabrea. She co-wrote a large paper, a thorough paper on the idea of the analytical engine, how it would be built and how it would be programmed. It is an amazing paper even not considering that it was written in the early 1840s. It's just a thoroughly amazing paper. You can find it on the web, it's well-worth reading to see how somebody who was just pioneering many of these ideas, who was thinking about these things for the first time really in human history was starting to think about the notion of automatic computing. In the course of that paper, Countess Ada says something very interesting and something that had been echoed or has been echoed since the electronic or digital computers have been made. Namely, she didn't quite use this language but this is essentially what she was saying, a computer can't do anything on its own, it can't think on its own, rather, it can only do those things that the programmer has told it to do. So it doesn't have any innovative intelligence. It just does the things that a program could do. That's a very interesting idea, and that's one of the arguments that Turing alludes to in discussing his own thought experiment of the Turing Test. There are different ways of responding to this objection. One, speaking as a programmer, I can tell you and you may have had this experience too if you're a programmer, speaking as a programmer, I can say that it is often the case that a program does things that I suppose in the sense I told it to do but were very unexpected to me. In other words, yeah, the program is doing what the programmer told it to do, but the programmer doesn't always know what he or she is telling the computer to do. Maybe more powerfully, you could certainly see the possibility of a computer program learning from experience over time. In which case, depending on the experience that the program has, it learns to do things that are well beyond anything that the programmer might have initially thought to put into the program. So these are potential responses to Countess Ada's very thoughtful and interesting objection. That's not the only reason to read Countess Ada's paper, it's a brilliant paper and has many other wonderful ideas and predictions in it as well. Finally, I should mention the objection from ESP, I don't have a slide for that but Turing was responding to experiments that had recently been done by a researcher named J B Rhine, R-H-I-N-E, who was working, I believe at Duke, and who had done a variety of these experiments, I'm not positive I'm getting this right, but it's something along the line like telepathy, where one person would look at a card with a symbol on it and another person across the room would try to guess what symbol the first person was looking at, and according to Rhine, there were people who were able to guess the observed symbol at a much higher rate than chance. I'm not going to go into arguments about these experiments here, suffice it to say that those experiments of Rhine have never been replicated, certainly not satisfactorily replicated. Frankly, I find the results very dubious. But if you're reading Turing's paper and you see him allude to this objection from ESP, you might think that he's being facetious but he's not, he's just keeping an open mind, like a good scientist and saying, "Well, if these experiments are on the level, then there may be a great deal more involved in thinking or intelligence than we've been ready to acknowledge this far." So he's just saying, if that's true, then perhaps there's something going on that we could not mimic with a computer. In point of fact, I don't think it's in retrospect, I don't think it's a very powerful objection. Well, what about the Turing test itself? Turing made a prediction at the end of his paper that he felt that the test would be passed by a computer program around the year 2000. As I'm speaking now, it is the year 2017 and no program has come close to passing the Turing test. There is no program out there such that you could do this kind of experiment, where you put the program behind one door and a person behind another door and the program could reliably fool a judge into thinking that it's a person after about 15 minutes or so. Nothing close to it. Which leaves us with a rather interesting question. If the Turing test cannot be passed, why? What was Turing missing? Maybe one of those objections that he was referring to is more powerful than he gave it credit for being. In other words, maybe there is a deep philosophical reason why no computer program could ever pass the Turing test. That's why, we, the community of computer scientists and researchers, have not managed to come up with a successful program in 67 years since Turing wrote this paper. As an aside, it's not that this has been the main goal of the computer science community in those intervening 67 years, people for a long time, there was a general lack of interest in actually trying to program a computer to do this kind of thing. You might say that, "Well, people have not really worked very hard on this problem, so maybe it's not that surprising that no program has passed the Turing test." Okay. There's some merit to that, and a lot of computer scientists, and researchers, and artificial intelligence do not feel that the Turing Test is a terribly meaningful test at least as far as their own research is concerned. But let's leave that aside and just say that, in fact, no computer program has passed the test, maybe it can't be passed. On the other hand, if it can be passed, look the other side of the question. If the Turing test can be passed, what are we missing? Why has it taken so long? So that's a kind of interesting pairing of questions. If the test can be passed, then what's taking us so long? If the test can't be passed, what is the reason for it? We don't have satisfactory answers to either of those questions. But we do have programs that in one fashion or another do pass the little local Turing tests, they're not general Turing tests of the kind that the paper imagines, but they're still interesting. The Watson program, the famously did so well on the Jeopardy quiz show, is one example, and you're probably familiar with a tiny little Turing tests that you see on the web all the time, the CAPTCHA program, which is, in essence, a little Turing test, you're now trying to pass the Turing test to show that you are a human being, not a computer. So this is a little test to see if you can read a couple of poorly written words, this would be a task that computer programs would have some difficulty doing, but people are able to do easily. There's a lot of interest even if you think about this kind of task. First of all, as you may have noticed, computers get better and better at passing these tests. So there's a arms race involved. Over time, the CAPTCHA questions have to become more challenging for the person to the point where, at least in my experience, sometimes I'm trying to get onto a web page and I get a question like this, and I don't think I can read the word that they're telling me to read. So it's getting to the point where I'm not a sufficiently good human being to actually pass the CAPTCHA test. But these are the kind of tasks that we'll also return to in discussion as well in this course. There are other kinds of small-scale Turing tests like whether a computer is able to write music, for instance. That might convince you that it was written by a human being. Other kinds of things like that, all quite interesting. Finally, I should not leave this topic without mentioning a very famous response to Turing's paper written I think around 1980 by the philosopher John Searle. He posed a counter thought experiment to Turing's test and his thought experiment is called the Chinese room experiment. The idea behind this experiment is you imagine, here's his scenario. You have a room. I'll just draw the boundaries of the room here. So, and there's the outside of the room and on this side on my side of the screen here, is the inside of the room, and there's a guy on the inside of the room and he has a large rule book. This is my highly imperfect drawing of a large rule book, and lots and lots and lots of scrap paper in a big basket. So lots and lots of scrap paper and a pencil, as much paper and as many pencils as he could possibly need. There is a little input slot in the room and a little output slot and here's the idea. Here is this guy. Let us assume that he, like me, is a native English speaker and does not speak a word of Chinese. Okay? This entire room is going to be playing the role of the computer in the Turing test. It's going to be answering questions in Chinese and the input slot therefore, has questions written in Chinese characters. The guy on the inside of the room does not understand a single symbol of Chinese. However, he's able to take these input characters and then use this rule book, which has a bunch of rules that tell him how to manipulate other kinds of characters, in response to these characters. He can write down any notes that the book tells him to write down on the scratch paper. So he has all the scratch paper he needs. He follows very explicitly to the letter the rules given in the rule book, and the rules are such that they tell him eventually to get another scrap of paper and write on it a certain number of characters, which are totally meaningless to him, to the guy, and then he passes those characters out the output slot. In other words, think of the entire behavior of the room. It gets Chinese questions in, this guy who does not understand Chinese is taking the rules in this book and using them to manipulate various marks on paper and eventually write some other Chinese characters, and presumably, the word book is written in English so he can understand the rules, but the rules tell him to write down certain squiggles and squaggles that he doesn't understand and then he passes those to him meaningless symbols out the output slot. Searle argues that, conceivably this room might pass the Turing test in Chinese if the program were good enough to mimic the behavior of a Chinese speaker. So Searle says, well, what would happen if, suppose this room did pass the Turing test, where is the knowledge of Chinese in here? What knows Chinese here? It's certainly not the guy, he doesn't know a word of Chinese. The book doesn't know Chinese. You can't say that a book knows anything. The book doesn't know Chinese. Certainly the scrap paper doesn't know Chinese. So what is it? Where is the knowledge of Chinese in this entire system? In this brief summary probably as far as Searle is concerned doesn't begin to do justice to his argument, but basically, the argument is that, you could have a mechanical entity that passes the Turing test, but it would not have what we would call consciousness. It's just a set of mechanical rules for manipulating symbols. Like Turing, Searle includes a variety of objections to his Chinese room scenario that he has heard from various people in presenting this thought-experiment. People that is responding to Searle's thought experiment and saying, no, I think you're presenting an unfair thought experiment to the purposes of artificial intelligence. So I would leave to you the task of reading Searle's argument, the very pleasurable task of reading slots paper, and going through the arguments that are made in response to the Chinese room, and thinking about his responses. There are many wonderful responses in the literature as well. People who have written since Searle's original paper, there have been many people who have responded to the Chinese room argument, and I have some favorites among those counter responses. But I won't go into them here and would just leave it to you to read Searle's argument. Let's leave it with this position. We have as yet unmet challenge from Alan Turing for machine intelligence. We have a provocative, but certainly not universally agreed to counter argument from John Searle saying that, even if a computer could pass the Turing test, and none have done so so far, but even if a computer could pass the Turing test, that would not be, for us, a very important or meaningful step in as much as this would not indicate that the program passing the Turing test has actual consciousness. What we would call meaningful Intelligence perhaps. So we'll leave it at that and as we continue to discuss Machines in mind, we'll talk about what things have been accomplished in computer programming to imitate human intelligence.
