Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!cornell!rochester!yamauchi From: yamauchi@cs.rochester.edu (Brian Yamauchi) Newsgroups: comp.ai Subject: Re: Question on Chinese Room Argument Message-ID: <1989Feb20.210414.9543@cs.rochester.edu> Date: 21 Feb 89 02:04:14 GMT References: <4298@pt.cs.cmu.edu> <51123@yale-celray.yale.UUCP> Reply-To: yamauchi@cs.rochester.edu (Brian Yamauchi) Organization: U of Rochester, CS Dept, Rochester, NY Lines: 107 In article harnad@elbereth.rutgers.edu (Stevan Harnad) writes: >[This is the negative note on which Searle's Argument ended in 1980; >not to leave it at that, let me add that in "Minds, Machines and >Searle" (1989) I've tried to take it further in a positive direction, >showing that it's only the symbolic approach to modeling the mind >that's vulnerable to Searle's Argument; nonsymbolic and hybrid >symbolic/nonsymbolic models are not. And in "Categorical Perception" >(1987) I have sketched how symbolic representations could be grounded >bottom-up in nonsymbolic (analog and categorical) representations. Now, >being immune to Searle's argument doesn't guarantee that a model has >captured understanding, of course (nor does it "effectively define" >understanding). But it does perhaps correct the misapprehension that >the validity of Searle's argument (and it IS valid) would entail that >NO model could understand; perhaps this misapprehension is behind the >strained, implausible and incoherent counterarguments people have tried >to float under the general banner of the "Systems Reply." You don't >have to give up on "systems". Just give up on purely symbolic systems.] >-- >Stevan Harnad INTERNET: harnad@confidence.princeton.edu harnad@princeton.edu I have been following this discussion for a while, and so I decided to go and read Searle's "Minds, Brains, and Programs" in Mind Design. In this essay, Searle outlines his basic argument and then tries to argue against a number of the possible objections. I think that Searle *does* have a valid criticism of traditional, symbolic AI. On the other hand, many of his counterarguments trying to broaden this point seem to range from the unclear to the bizarre. The basic idea that symbol manipulation alone is not necessary for intelligence makes sense. To translate Searle's argument from the language of philosophy to the language of AI, consider what it means to understand the sentence "The dog chased the cat." Traditional AI would represent this as: dog(x) & cat(y) & chased(x,y) However, the program really has no idea what a dog is, what a cat is, or what it means from one thing to chase another. This, I believe, is the crux of Searle's argument. One could add a dog schema which said something like: dog is-a : animal (subtype : mammal, carnivore) environment : land legs : 4 tail : yes But, then the program still doesn't know what a land environment is, or what legs are, etc. The conventional counterargument is that the richness of the knowledge base determines the level of understanding. So one could add schemas for mammals and environments, and so forth. Of course, these would have to be defined in terms of other symbols, and so on and so forth. To a large extent this is what happens with human learning. We learn new concepts by relating them to things we already know. The *critical* difference, in my opinion, is that at some level all of our learned symbols are grounded in sensory experience. Most of us probably learned what a dog was by seeing one or by seeing a picture of one, not by reading a dictionary definition. We know that "chasing" refers to an activity that we have seen (on TV, at least, if not in person), rather than simply a construct of: chase(x,y) --> wants-to-catch(x,y) & wants-to-avoid(y,x) Therefore, in order to build a system that "understands" in the same way that people "understand", we need to give it the ability to relate the concepts in its knowledge base to sensory experiences. This is a similar to what Searle calls "The Combination Reply" -- that a complete robotic system with sensory perceptions and motor control (and possibly based upon neural networks) could be said to have "understanding". Searle admits (p. 296) : "I entirely agree that in such a case we would find it rational and indeed irresistable to accept the hypothesis that the robot had intentionality, as long as we knew nothing more about it." But then, he goes off and says that since we know how the robot works, we can't ascribe "intentionality" to it. He says we *can* ascribe "intentionality" to animals because (1) We don't understand how they work and (2) They are made out of the same stuff as humans. This is almost too absurd to contemplate. (1) is equivalent to arguing that since primitive man couldn't explain the weather without refering to magic, storms must be the result of sorcery. (2) is nothing more than a form of vitalism which might be understable if Searle were a mystic, but is all the more baffling since he states (p. 300) the materialist position that humans are, in fact, machines that think. Searle goes so far as to state "Whatever else 'intentionality' is, it is a biological phenomenon and it is as likely to be causally dependent on the specific biochemistry of its origins as lactation, photosynthesis, or any other biological phenomena." One can only wonder what would happen if it were discovered that some humans depend more heavily on some neurotransmitters than others. Who would Searle consider non-intentional: the people using the non-standard neurotransmitters or whoever was using neurotransmitters that were different from his own? _______________________________________________________________________________ Brian Yamauchi University of Rochester yamauchi@cs.rochester.edu Computer Science Department _______________________________________________________________________________