jufier@daimi.aau.dk (Svend Jules Fjerdingstad) (11/12/90)
The January 1990 issue of Scientific American featured two articles about Artificial Intelligence: "Is the Brain's Mind a Computer Program?" by John R. Searle, and "Could a Machine Think?" by Paul M. Churchland and Patricia Smith Churchland. All three authors are professors of philosophy, which may explain their poor general understanding of the properties of computers and computer programs. In the September issue Scientific American printed a number of responses to the articles. Although I agree with most of the objections to the articles stated in these letters, I feel that important points of criticism was omitted. The following represent my response to the articles. I must confess, that I'm not exactly impressed by the quality of the arguments presented in the two articles. It seems to me, that most of the arguments are quite weak and remarkably simple to refute. In "Is the Brain's Mind a Computer Program?" professor of philosophy John R. Searle attempts to show that even passing the famous Turing test does not prove a computer program intelligent. He tells this story of a "Chinese Room" in which a person ignorant of Chinese language manipulates Chinese symbols according to the rules in a book. He correctly points out, that the rule book is the "computer program", and that the person is the "computer", merely executing the program. But then he wrongly concludes, that the _person_ satisfies the Turing test in spite of being ignorant of Chinese. Obviously, it is the Chinese Room _as a whole_, which is able to pass the Turing test. And the room itself certainly cannot be said to be ignorant of Chinese, as an extensive knowledge must be present in the rule book, although the person in the room has no means of accessing this knowledge. To him or her, the rules in the book _seem_ entirely meaningless. Unless speaking Chinese in a manner "indistinguishable from those of a native Chinese speaker" does not require intelligence, then of course the entire room must be considered intelligent. It certainly behaves intelligently, and as this cannot be ascribed to the person, the intelligence must be due to the rule book, whether the rules are stored on paper or not. Searle tries to avoid this so-called Systems Reply by imagining yet another (impossible) situation, in which the person in the room memorizes all the rules for manipulating the symbols. His argument is nearly unbelievably naive: "There is nothing in the 'system' that is not in me, and since I don't understand Chinese, neither does the system." The consequences of this statement are absurd: If his statement is correct, then he has proven nothing more, than that it should be possible for a person ignorant of Chinese to pass the Turing test for speaking Chinese. All he has done is reducing the requirements for speaking a language fluently. In fact, he has simply made it impossible to determine, whether a person understands a given human language or not: If other humans are able to answer in speech as though they understood what was being said, then surely all their actions might also be the result of consulting a simple rule book. The entire concept of comprehension has been made hollow and without meaning. Furthermore, if the Turing test no longer is a valid test for intelligence, then it has become impossible to judge the degree of intelligence present in a person. The concept of intelligence has also been made meaningless, as the presence or absence of Searle's kind of intelligence in no way influences an entity's behaviour. If it did, the Turing test could be used to distinguish between the two. Of what use are Searle's concepts of comprehension and intelligence, if they are not related to any events of this world? If Searle's statement were true, he would simply have made it clear that conscious thought was unnecessary for all kinds of human behaviour! Which is probably the exact opposite of what he wanted to prove. But perhaps the key to the explanation could be, that Searle himself actually does not understand English. Whenever he writes an article, he simply consults a (lousy) book full of rules for writing articles against artificial intelligence :-) The entire attitude of Searle is ridiculous. He states that "a program merely manipulates symbols, whereas a brain attaches meaning to them". But why is it important, that humans attach meaning to symbols? Does it matter? Of course it does, it is exactly this attaching meaning to symbols that allows us to interpret the symbols as conveying a message, and allows us to understand this message. Obviously, the fact that we are able to understand symbolically encoded messages affect the way we interact with our environment. Which is why a computer would have to be able to understand human language, not merely manipulate it, in order to pass the Turing test. Does our intelligence influence the way we behave? Of course, it does. If not, what is intelligence, then? And why are we equipped with it, if it is of no real, practical use? But if intelligence influences our behaviour, then it follows, that a computer also would have to be intelligent, if ever to pass the Turing test. As the person in the Chinese Room were unable to understand the Chinese signs entering the room, all of these qualities would have to be present in the "rule book" in the Chinese Room, which would make it something far more sophisticated than merely a book. I think, this is where Searle really cheats. By using a simple book full of rules, he is able to carry along with his arguments, because it is obvious, that a book could not understand Chinese, or be intelligent. However, he has at no time demonstrated (or even made likely) that a book would indeed be sufficient for his project. His assumption is, in fact, a degradation of intelligence to a set of simple rules that can easily be formulated in a book. Searle gets things mixed up by assigning the name "book" to an entity with properties, that are fundamentally different from those of a book. This is, in my opinion, the central error in the Chinese Room argument. Searle confuses himself (and the Churchlands) by calling something a book, that could never be just a book of rules. Searle's attack on the Turing test is unfair and unfounded. He mistakes his own lack of understanding for flaws in the Turing test. It is odd, that the Churchlands in their article "Could a Machine Think?" fail to recognize the untenability of Searle's arguments against the Turing test. They agree with Searle, that "it is also very important how the input-output function is achieved; it is important that the right sorts of things be going on inside the artificial machine." Well, I certainly do not agree with that. That's just pure mysticism. A black box must be considered intelligent, if it acts intelligently. It is ridiculous to define "conscious intelligence" in a way, which makes it impossible to measure, because it in no ways affect its surroundings. With this definition, we won't ever be able to determine whether conscious intelligence is present or not in an object. All we can say is the following: "Is it human? Ah, then it's intelligent! Not human? Well, then this seemingly intelligent behaviour is not achieved the right (human) way. Therefore, it is not intelligent!" Then we have defined intelligence in such a way, that it can only occur in humans. And by defining it this way, we have excluded all the most impressive and important qualities of human intelligence. I am not at all impressed by these three philosophers' abilities to reason: Searle also argues that simulating a process on a computer is very different from the actual process, and that therefore even if we could simulate all the processes of the brain, we should still not have attained artificial intelligence. However his example, that you cannot make a car run with a computer simulation of the oxidation of hydrocarbons in an engine, is really not relevant. Obviously, "a simulation of cognition" would process the same genuine information as a brain, not just some symbolic substitute. And it would process it the same way as the brain. The simulation would be a genuine "processor of information", just like the brain. Therefore the two situations are not comparable, and the argument is invalid. Let me try to summarize, why I think AI is possible. The definition of intelligence I will use, is the more usual one than Searle's: Intelligence is the ability to interact with the environment in an "intelligent" way, that is, in a way which shows comprehension of the workings of the environment. My argument goes like this: Only purely random events (if such exist at all) are not governed by rules. And since intelligence is the quintessence of non-randomness, rules for intelligent behaviour must exist, however complex they may be. These rules are not the kind of rules to be found in Searle's rule book. These are complex rules, which take into account all knowledge and memory of past experiences, all emotions, the behaviour of the surroundings, et cetera. Because of that, all rules are not the same in all humans, but no doubt we share a large proportion of these rules. Intelligent behaviour is in essence highly non-predictable. But this is simple due to the complexity and multitude of the rules guiding intelligence. It is certainly not the result of no rules, as this would only lead to random, and thereby non-intelligent, behaviour. As a consequence of the existence of such complex rules, artificial intelligence is possible, as it is "simply" a matter of creating a machine, which is able to handle these complex rules and the enormous amount of memory required. And it would be very unlike human beings, if we were not to achieve such a machine one day. In his response in the September issue of Scientific American, Searle writes: "It is a mistake to suppose that in opposing a computational explanation of consciousness I am opposing a mechanical explanation." In my opinion, this doesn't make sense at all. If there is a mechanical explanation, then there has to be a computational explanation as well, because every mechanical process can be described computationally. Searle admits: "There must be a mechanical explanation for how the brain processes produce consciousness because the brain is a system governed by the laws of physics." Yes, and that is precisely the reason why artificial intelligence is possible. "Consciousness" and "intelligence" are results of a functioning brain. It is this functionality, that we want to recreate. What matters is the functionality itself, not how it is achieved. And if there is a mechanical explanation of how the brain's information processing works, then of course we can, in principle, recreate it as a computer program, however complex it might have to be. Searle ends his response my stating that: "Any sane analysis has to grant that the person in the Chinese room does not understand Chinese." Right, I agree. I don't think, that Searle could find anybody disagreeing. However, this is completely uninteresting. As Searle pointed out in his article, the person is acting as a "computer", whereas it is the rule book, that corresponds to the computer _program_. So Searle has just proved, that the _hardware_ of a computer need not be intelligent in order for the computer _system_ to be intelligent. But he certainly hasn't proved, that the brain's mind could not be the result of executing a computer program, as was his intention. Searle's article, and others like it, always makes me think of the apparent paradox, that the people most strongly opposed to the notion of artificial intelligence are sometimes those, who seem less well endowed with natural intelligence :-) -- Svend Jules Fjerdingstad, jufier@daimi.aau.dk Computer Science Department, University of Aarhus Ny Munkegade 116, DK-8000 Aarhus C, DENMARK
deichman@cod.NOSC.MIL (Shane D. Deichman) (11/14/90)
In his earlier posting, Svend makes some brilliant arguments in support of a deterministic, non-free will environment for human existence. By deftly casting the arguments of both Searle and the Churchlands aside, he resorts to a "If it exhibits the qualities of intelligence then it IS intelligent" argument. Is that to say that human perceptions are always infallible, and that what we see and perceive actually IS? Or does it imply that our percep- tions, while not always accurate, still elicit a deeper understanding of a given phenomenon based on multiple repetitions? The Chinese Room argument points out some deficiencies in the Turing Test -- deficiencies which call upon the observer to take a deeper, more profound look at what is meant by "understanding" and "knowledge." Svend disregards the subconsciousness associated with cognition and lucidity, and therefore begs the question. Furthermore, he attacks the Churchlands (supposed "allies" in his campaign in support of Strong AI) in their reasoning capacities for failing to see this point he so astutely raises. Perhaps, in a stolid, deterministic world where emotions are bleak representa- tions of mere "sensory inputs," Svend's arguments would carry some weight. But in a world enriched by the subtleties of life, his "intelligence" as a function of outward appearance is exceedingly bland. -shane "the Ayatollah of Rock-and-Rollah"
JAHAYES@MIAMIU.BITNET (Josh Hayes) (11/14/90)
What needs defining here is "intelligence", because it seems that Searle has his own definition which _de facto_ includes being a human being, or at least an organic being; it's no surprise then that no machine "intelligence" need apply.... Sven, on the other hand (I have a Colombian friend named Sven; it's an odd name for that part of the world....where was I?) defines intelligence as "that which appears intelligent" (I paraphrase, but I think, fairly). This is a simple definition (though it begs the question of how we determine what "appearing intelligent" is), and is, I think, the relevant definition to the question of A.I. I believe we want a pragmatic definition: what is the PURPOSE of AI? We ostensibly design AIs to perform a task or tasks that we assume to require a degree of intelligence; to the extent that they carry them out well, are they not intelligent? This all ties back to the emergent properties shtick (sorry). The systems reply to Searle's CR analogy is entirely appropriate if we regard "intelligence" as a property of a system as a whole which cannot be said to reside in any particular component of that system. It is the property of the "instruction book" and the "guy who manipulates the symbols" AND the interaction between these sub- systems AND the interaction of that whole system with the outside world (that speaks Chinese to the "Room"). As such, "intelligence" may not be a very useful term, since it's so difficult to pin down, and of course, since it's such a loaded term. ------- Josh Hayes, Zoology Department, Miami University, Oxford OH 45056 voice: 513-529-1679 fax: 513-529-6900 jahayes@miamiu.bitnet, or jahayes@miamiu.acs.muohio.edu "It is always wise to remember that it was the gods who put nipples on men, seeds in pomegranates, and priests in temples."
jufier@daimi.aau.dk (Svend Jules Fjerdingstad) (11/17/90)
deichman@cod.NOSC.MIL (Shane D. Deichman) writes: >In his earlier posting, Svend makes some brilliant arguments in >support of a deterministic, non-free will environment for human >existence. By deftly casting the arguments of both Searle and >the Churchlands aside, he resorts to a "If it exhibits the qualities >of intelligence then it IS intelligent" argument. Is that to say >that human perceptions are always infallible, and that what we >see and perceive actually IS? Or does it imply that our percep- >tions, while not always accurate, still elicit a deeper understanding >of a given phenomenon based on multiple repetitions? No. The point is this: If a human being "exhibits the qualities of intelligence" (according to our (subjective) perception of such qualities), then we DO (in normal every-day life) consider this human being to be intelligent. Therefore if some entity (be it a computer system or anything else) behaves "intelligently", then we MUST also conclude, that this entity has intelligence. If we cannot consider a computer system intelligent EVEN THOUGH it behaves intelligently, then we have redefined the concept of intelligence in such a way as to make it completely unrelated to any behaviour, that we can observe. This means that any piece of dirt might indeed be considered intelligent, or alternatively, that it is impossible to conclude about any human being, that he or she is intelligent. This definition could, in fact, lead to a belief in the non-existence of true intelligence, whether in humans or in computers. (Except in me, of course :-)) In my opinion, this last definition of intelligence is absurd an useless. Intelligence is the ability to BEHAVE intelligently. Nice definition, eh :-) The problem is that we cannot at the present time (and perhaps we never will be able to) give a precise and exhaustive definition of intelligent behaviour. Therefore the Turing test represents the brilliant solution of using one intelligent system, human beings, to evaluate the possible degree of (verbal) intelligence residing in some other supposedly intelligent system, a computer system, for example. >The Chinese Room argument points out some deficiencies in the Turing >Test -- deficiencies which call upon the observer to take a deeper, >more profound look at what is meant by "understanding" and "knowledge." >Svend disregards the subconsciousness associated with cognition and >lucidity, and therefore begs the question. If subconsciousness is a prerequisite for intelligence, if it plays a role in forming intelligent behaviour, then of course a computer system would have to possess subconsciousness, in order to pass the Turing test. Anyway, IMHO the only deficiencies pointed out by The Chinese Room argument are deficiencies in Searle's understanding of the Turing test :-) If Searle's Chinese Room argument were valid, then all of you people out there on the net might just be mindless machines looking up words in a dictionary. But then, why do I bother writing this? Better stop now :-) >Furthermore, he attacks the Churchlands (supposed "allies" in his >campaign in support of Strong AI) in their reasoning capacities >for failing to see this point he so astutely raises. Perhaps, in >a stolid, deterministic world where emotions are bleak representa- >tions of mere "sensory inputs," Svend's arguments would carry some >weight. But in a world enriched by the subtleties of life, his >"intelligence" as a function of outward appearance is exceedingly >bland. Ah, I thought so. You ARE one of those :-) >-shane >"the Ayatollah of Rock-and-Rollah" Svend -- Svend Jules Fjerdingstad, jufier@daimi.aau.dk | "To love, Computer Science Department, University of Aarhus | and to learn." Ny Munkegade 116, DK-8000 Aarhus C, DENMARK |
marky@caen.engin.umich.edu (Mark Anthony Young) (11/18/90)
In article <1990Nov16.161134.2845@daimi.aau.dk> jufier@daimi.aau.dk (Svend Jules Fjerdingstad) writes: >deichman@cod.NOSC.MIL (Shane D. Deichman) writes: > >>The Chinese Room argument points out some deficiencies in the Turing >>Test -- deficiencies which call upon the observer to take a deeper, >>more profound look at what is meant by "understanding" and "knowledge." >>Svend disregards the subconsciousness associated with cognition and >>lucidity, and therefore begs the question. > >If subconsciousness is a prerequisite for intelligence, if it plays a role >in forming intelligent behaviour, then of course a computer system would have >to possess subconsciousness, in order to pass the Turing test. > I think this is a very important point, one that is ignored in the Chinese room argument. The CR argument goes like this: IF the Turing test is correct, AND a machine of such-and-such a type passes it, THEN that machine is intelligent. BUT, that type of machine can't be intelligent (it doesn't have the "right stuff") THEREFORE, the Turing test is not correct. Implicit in this argument is that the offending machine will pass the Turing test (otherwise the implication is invalid). It is possible that the Turing test is valid, and yet no machine will ever pass it. It is possible that any machine that passes the TT will be totally unlike anything we now consider to be a computer. Nevertheless, this would not invalidate the test itself. ...mark young