smoliar@vaxa.isi.edu (Stephen Smoliar) (01/15/91)
In article <26250@uflorida.cis.ufl.EDU> bougie@pine.circa.ufl.edu writes: > Basically Reddy shows how we use a metaphor >to talk about language which doesn't fit the phenomenon very well and >leads us into a lot of false analogies. We speak about "finding meaning >*in* words, packing too much/little meaning into a sentence, getting >meaning *out of* a phrase..." etc etc (the appendix is impressive). We >tend to think of language as little boxes that we fill with meaning, and >send down a conduit to a receiver who then unpacks the boxes. This >leads to the assumption that if I don't find any meaning in the box, >it can only be the fault of the sender! > I think this metaphor (along with its contingent dangers) may be readily extended from the concept of "language" to that of "knowledge." (This discussion may actually be more appropriate on comp.ai.philosophy, so I am cross-posting this article.) There seems to be an underlying theme in Newell's Knowledge Level Hypothesis (which is expanded upon at some length in Pylyshyn's COMPUTATION AND COGNITION) that knowledge is some kind of "stuff" which we can use to fill "vessels" of some sort or transfer from one vessel to another using a Reddy-like conduit. Ultimately, Newell and Pylyshyn (not to mention others, such as Fodor) argue that it either IS or, in a weaker case, CAN BE MODELED BY symbolic expressions. However, what if this whole "stuff" metaphor is as misplaced for knowledge as it is for language? This would knock a fair amount of life out of Newell's Physical Symbol System Hypothesis and all that follows from it (such as the Knowledge Level Hypothesis). > Reddy sees this as not only mistaken, but harmful. People often >find it difficult to talk about language at all without using the >Conduit Metaphor. He proposes the "Toolmaker's metaphor" as a better >description: more like sending a *blueprint* for reconstructing >meaning, than sending *meaning* itself. > > Thus, the bulk of the *message* is not *sent*, but constructed by >the hearer from relatively very few bits that are actually sent along >the conduit. Inferences make up an enormous part of the meaning. > However, if we try to think about inferences in terms of a logical calculus, we are back to the same symbolic "stuff" I am trying to get away from! Even connectionism, while some would stand it in opposition to the Physical Symbol System Hypothesis, may ultimately be reduced to some sort of "stuff-like" representation, where the "stuff" is now points of convergence, rather than symbolic expressions. In other words connectionism may be able to transcend the symbols without escaping the "stuff" metaphor. The only work I know which has tried to pull away from this metaphor is that of Gerald Edelman and his colleagues. The automata which Edelman has tried to design treat perceptual categorization as a lowest-level task to be achieved. However, Edelman's categories are not static entities, corresponding to the sorts of local maxima sought out by a connectionist architecture. Instead, they are far more dynamic. Memory is not a matter of accumulating more "stuff." Rather, it is a capacity for categorization and REcategorization, the latter being his way of expressing the sort of processing involved when confronted with stimuli one has encountered before. This is very much in line with the sort of constructive metaphor proposed by Reddy; but Edelman carries it to a much greater extreme, ultimately arguing that it lies at the heart of all cognitive activity. ========================================================================= USPS: Stephen Smoliar 5000 Centinela Avenue #129 Los Angeles, California 90066 Internet: smoliar@vaxa.isi.edu "It's only words . . . unless they're true."--David Mamet
jmc@DEC-Lite.Stanford.EDU (John McCarthy) (01/15/91)
I think that writing about AI or philosophy or cognitive science in terms of metaphor is a big mistake. It allows people to write without clear meaning. The debates about which metaphors are applicable are almost meaningless. There are several approaches to AI, some based on neurophysiology, some on imitation neurophysiology, some on psychological experimentation, some on formalizing the facts about the common sense world in mathematical logical languages (my own course). There is no argument that any one of them can't possibly work. Therefore, AI research is a race among the various approaches. The arguments about metaphor are a game for non-participants in the actual work. Buy my book.
ggm@brolga.cc.uq.oz.au (George Michaelson) (01/15/91)
[re-blocked to suit interpolated comments] jmc@DEC-Lite.Stanford.EDU (John McCarthy) writes: >There is no argument that any one of them can't possibly work. ...You mean AI workers don't disagree about the relative merits of their model in pejorative terms? amazing! Outside of the field, I suspect scepticism remains that ANY of them can possibly work. >Therefore, AI research is a race among the various approaches. - a Red Queens race perhaps? >The arguments about metaphor are a game for >non-participants in the actual work. In my own case, undenyably true! I think they also point to the weakness of available models. If nothing else, its an overspill of ideas from the lofty heights of the castle to the rude huts of the commoners below. When the metaphors start becoming testable and/or (dis)provable theorems, then things will be a little more solid perhaps. If you're asserting that behind this peat-bog of metaphors lies a more solid ground of theory I'll sink back into the mud from whence I came. Like all creatures of the (CS) slime, I tend to remain skeptical of these (AI) attempts to walk on solid land. -George -- G.Michaelson Internet: G.Michaelson@cc.uq.oz.au Phone: +61 7 365 4079 Postal: George Michaelson, the Prentice Centre, The University of Queensland, St Lucia, QLD Australia 4072.
smoliar@vaxa.isi.edu (Stephen Smoliar) (01/16/91)
In article <JMC.91Jan14145958@DEC-Lite.Stanford.EDU> jmc@DEC-Lite.Stanford.EDU (John McCarthy) writes: > >Buy my book. Wouldn't you be happier if we were to READ your book, John? (Man does not live by royalties alone!) In any event, my current "addiction" to metaphor will not dull my curiosity about what you have to say. Would you be kind enough to provide us all with publication details? ========================================================================= USPS: Stephen Smoliar 5000 Centinela Avenue #129 Los Angeles, California 90066 Internet: smoliar@vaxa.isi.edu "It's only words . . . unless they're true."--David Mamet
jmc@DEC-Lite.Stanford.EDU (John McCarthy) (01/16/91)
I never thought you'd ask. Formalizing Common Sense, Ablex 1990. (actually 1991).