kck@g.gp.cs.cmu.edu (Karl Kluge) (03/21/89)
I tried to take this to email, but the path addrmap gave me for Roel W. didn't work. Hopefully the subject line will get caught by people's kill files if they don't want to follow this.... It's unclear to me what question or objection R. W. is making. The first post in the series says > How then could a universal TM (i.e. a computer) fed with a program which can > answer questions in Chinese ever come to "know" the denotation of the > symbols it is manipulating? The outcome of its computation is invariant > under changes of denotation of the symbols it manipulates; the people > programming the UTM may change the denotation of symbol xyz from chair to > table or to anything else, without it making the slightest difference to the > computation. What do you mean by "knowing the denotations of the symbols it manipulates", and why do you consider this relevent to the question of whether a TM can "think"? If "know" really means "know" here, then I'll have to confess that neither I nor anyone in AI has an answer. People have been arguing over epistemology for 2500+ years, and we still can't prove that *we* "know" anything, let alone whether a Turing machine can "know" something. This isn't to trash philosophy, just to acknowledge its limitations. If what you meant to ask is how a TM can represent the denotation of its symbols, then (if I'm not mistaken -- I may well be, as formal logic isn't my area) this is the symbol-grounding problem. Several answers have been offered (similar to Steve Harnad's), usually based on the correlation between the distal stimuli that are the denoted objects and the proximal stimuli that are generated in the sensors of a robot driven by a TM. Your reply, however, seems to indicate that this concerns the operational semantics of the symbols rather than the denotations. I still find the statement "the people programming the UTM may change the denotation of symbol(s)..." somewhat odd. If the symbol is an input or output symbol (let's restrict ourselves to a TM with text I/O for a second), then the denotations of these "public" symbols are not up to the programmers (as they are socially approximately agreed upon). If the symbol is a "private" internal symbol, then I fail to see that the programmers are in a priviledged position to "decide" what that symbol's denotation is. You seem to accept that when you say > However, for > symbol-manipulating system like a computer, I agree with the following: > > >There are no social conventions, and > >hence no privileged denotations for the system's "private" symbols. ********************* > What you confuse is knowing a denotation of a symbol and agreeing with other > people about what the denotation of the symbol is (during the discourse). I think many people (Deconstructionists come to mind) would deny the possibility of anyone "knowing" the denotations of the words they use other than by the apparent agreement of other people in discourse. Once again, I sense that the word you want to use is "represent" rather than "know". Is this the case? Part of the problem is that we have technical words like "represent" that we don't necessarily want to apply to minds, while we have natural language terms like "know" that we aren't sure it makes sense to apply to formal systems. Further, you seem to be happy punting the issue of whether people "know" the denotations of the symbols they use (regardless of whether people are "symbols all the way down"), but seem to feel that TMs "knowing" the denotations of their symbols is relevent to whether a TM can think. If we can agree that people can "think" without "knowing" the denotations of the symbols they use (regardless of whether people are "symbols all the way down"), then why is the ability of a TM to "know" the denotations of its symbols relevent to the question of whether it can think? > After showing that symbols can be realized physically, and that their > operational semantics can be realized physically in a causal process, you > remark that > > >The denotations of the symbols have no corresponding reality. > > Correct, but I don't see what that has to do with the problem of whether a > symbol-manipulation process can think. Unless, of course, you assume that > our brains realize thinking by realizing a symbol-manipulating process. I'm trying not to assume anything about the brain for a momment. The point I was trying to make is that "denotation" seems to be an arbitrary property of the symbols in an instantiated, executing formal system, unlike the syntax and operational semantics of the symbols. I don't understand your concern about the system's "knowing" this arbitrary property, when it can (by hypothesis) do all the right things in terms of interacting with the world without "knowing"/representing this property. Put differently, there are two questions we'd like to answer 1) Can a TM (attached to appropriate I/O devices) pass appropriate tests (the LTT or TTT or pick your favorite) that would make us comfortable saying that the TM could simulate a mind (an empirical question), and 2) Would such a machine actually have a mind (a philosophical question whose answer is not open to verification or disproof, since the answer you get will depend on the assumptions you start with). As a scientist, I find (1) interesting. As a human, I find (2) interesting, but irrelevent to my work. I don't understand why you feel that whether a TM can "know" the denotations of its symbols is related to the answer to (2). Searle purports to have proven that the answer to (2) is no. I don't buy his argument because I don't believe that the truth of what he calls "Strong AI" would imply that he ought, while simulating a TM that understands Chinese, understand Chinese in the way he understands English. Remember, not all of what is going on in Searle's mind is causally directed by the TM he is simulating, so the presence of extraneous thoughts (like "I don't understand Chinese") in his mind (serially or in parallel with the simulation) doesn't necessarily tell us anything. No one defending Searle has chosen to address this problem. As an afterthought, Steve Harnad offered a definition of symbolic AI. I'd like to offer John Haugland's definition (he calls it GOFAI, Good 'Ole Fashion AI). This is taken from my notes on a talk he gave ************************************************************** a) Thinking is symbolic i) Symbols have semantic content which pertain to aspects of the situation ii) Symbols have a syntax iii) Meanings of atomic symbols are essentially arbitrary b) Process of thinking is regular or inference-like i) Rules expressed in terms of syntax of composition of symbols. Rules need not be explicit anywhere. ii) Rules suffice to constrain manipulation in ways that keep thought rational, have it make sense (example: valid inference rules). iii) Need not be explicitly coded. c) Thinking can be mechanized i) Rules built into causal structure of the physical machine. ii & iii) Missing in notes iv) Note that all that is required is that the machine be structured to carry out the manipulation -- medium independent (as opposed to Searle's "causal powers" argument). ************************************************************** The issue of rule-based vs. rule-following has been brought up. The classic example is "The path of the planets can be described by a set of partial differential equations, but the planets don't solve partial differntial equations as they orbit the Sun." Isn't the same point just as true of a digital computer (which, after all, is just an analog electrical circuit at bottom)? Suppose that I want to solve the set of diffential equations that describe the planets' paths, and choose to use the real planets as an analog computer to do it -- wouldn't we feel comfortable then saying that the planets were solving the equations in the same way that we say the analog electical circuit is executing a set of rules? Regarding the Hawking analogy -- there are two (related) questions 1) What kinds of interaction with the world does it take to develop a mind in an obect capable of having one (feral children, while possessing fully human brains, generally never develop fully human cognitive abilities), and 2) What kinds of interaction with the world do we want to see in order to be convinced that an object has a mind. On the output side, it seems to be the case (for instance, in children born with serious Cerebral Palsy) that someone who can make only a small, discrete (often binary) set of controlled movements can develop normal cognitive abilities, and can demonstrate that they have minds. Therefore, I don't see that the ability to perform the rich motor tasks a "normal" (I can't think of a more PC term right now, I apologize for any offense) person can engage in is relevent to developing/proving one has a mind. On the input side, Helen Keller provides an upper bound on how much one can reduce human sensory bandwidth and develop a "normal" mind. The lower bound is unknown. Karl Kluge (kck@g.cs.cmu.edu) "There is no practical difference between a very, very busy server and a dead server. They're like advisors that way." -- John Ousterhout --
lee@uhccux.uhcc.hawaii.edu (Greg Lee) (03/22/89)
From article <4532@pt.cs.cmu.edu>, by kck@g.gp.cs.cmu.edu (Karl Kluge): " ... " Regarding the Hawking analogy -- there are two (related) questions " 1) What kinds of interaction with the world does it take to develop a mind " in an obect capable of having one (feral children, while possessing fully " human brains, generally never develop fully human cognitive abilities), and " 2) What kinds of interaction with the world do we want to see in order to be " convinced that an object has a mind. " " On the output side, it seems to be the case (for instance, in children born " with serious Cerebral Palsy) that someone who can make only a small, " discrete (often binary) set of controlled movements can develop normal " cognitive abilities, and can demonstrate that they have minds. Therefore, I " don't see that the ability to perform the rich motor tasks a "normal" (I " can't think of a more PC term right now, I apologize for any offense) person " can engage in is relevent to developing/proving one has a mind. Yes. So while we're parsing questions, let's separate: 1) The evolutionary influence of motor and sense mechanisms on human thinking abilities, and 2) the current-day connection between motor and sense mechanisms and human thinking abilities. When we stick to normal cases, it's easy to mistake 1) for 2). Some speculation about the evolution of human language connects it with the development of a tongue and other articulatory facilities that can make a rich variety of sounds in a controlled way. That's as may be. But some people are born without tongues. They learn to understand and speak language. They don't speak well, but they can do it. I'm not sure just what Harnad's continual references to "transducer and effector surfaces" mean, but if the general idea is that human thought/understanding is somehow crucially dependent on human sense and motor mechanisms, though this is a plausible speculation, I think the evidence is against it as a synchronic hypothesis. Greg, lee@uhccux.uhcc.hawaii.edu