dhw@itivax.iti.org (David H. West) (02/28/90)
In article <12038@venera.isi.edu> smoliar@vaxa.isi.edu.UUCP (Stephen Smoliar) writes: | One of the major punch lines of Gerald Edelman's NEURAL |DARWINISM is that algorithms based on selection from a population may |allow us to deal with the first half of this question--category formation. I suspect that this constitutes a "punch line" only for those unaware of the work of John Holland. |Supposedly, he is taken on the assignment and manipulation of labels (which |brings us into the realm of symbols) in his new book, THE REMEMBERED SELF; See the chapter on the "broadcast language" in Holland's 1975 book. -David West dhw@iti.org
smoliar@vaxa.isi.edu (Stephen Smoliar) (03/01/90)
In article <5028@itivax.iti.org> dhw@itivax.UUCP (David H. West) writes: >In article <12038@venera.isi.edu> smoliar@vaxa.isi.edu.UUCP (Stephen Smoliar) writes: >| One of the major punch lines of Gerald Edelman's NEURAL >|DARWINISM is that algorithms based on selection from a population may >|allow us to deal with the first half of this question--category formation. > >I suspect that this constitutes a "punch line" only for those >unaware of the work of John Holland. > My personal feeling is that Holland's major advance has been in the area of ASSIGNMENT to categories, rather than the FORMATION of those categories. Thus, the selective capabilities of his genetic algorithms are good for determining how an input stimulus should be classified; but the categories themselves are essentially "hard-wired" into the architecture of the systems "genes." However, if I am mistake in this impression, I would be very interested in having someone set the record straight. By the way, Edelman and Holland are quite aware of each other, as Holland sat on the IJCAI panel which discussed Edelman's work. Much of the discussion involved the question of how an agent can manage in an unlabeled world, this being Edelman's point of departure for his category formation work. Holland's position statement argued that retinal patterns are just as much "labels" as are nouns in a natural language. It sounds to me as if this argument is trying to say that category formation is not an issue, and I'm not sure I'm ready to buy into that yet. I always liked Tony Hoare's observation that "the most powerful tool available to the human intellect is abstraction." I view nouns as an abstraction of retinal patterns (among other things); and I am not sure there is much to be gained from calling things labels which really do not impose ANY abstraction on the input stimuli. Edelman seems to be saying that, not only do such abstractions exist (i.e. abstractions on "raw data" like retinal stimuli), but also they have a biological foundation. ========================================================================= USPS: Stephen Smoliar USC Information Sciences Institute 4676 Admiralty Way Suite 1001 Marina del Rey, California 90292-6695 Internet: smoliar@vaxa.isi.edu "Only a schoolteacher innocent of how literature is made could have written such a line."--Gore Vidal
dhw@itivax.iti.org (David H. West) (03/03/90)
[a previous version of this was posted at 1am, when I don't think as clearly, but it couldn't be cancelled until morning because of local news peculiarities. Apologies to anyone who sees both.] In article <12084@venera.isi.edu> smoliar@vaxa.isi.edu.UUCP (Stephen Smoliar) writes: |My personal feeling is that Holland's major advance has been in the area of |ASSIGNMENT to categories, rather than the FORMATION of those categories. Thus, |the selective capabilities of his genetic algorithms are good for determining |how an input stimulus should be classified; but the categories themselves are |essentially "hard-wired" into the architecture of the systems "genes." |However, if I am mistake in this impression, I would be very interested |in having someone set the record straight. My personal view (nothing as definitive as "setting the record straight") is that categories are built from sub-categories; at the bottom, one (usually) hits hardware, and above that, one can build as many evolving layers as one has the leisure to wait for. People who have to get results out of serial simulations before their funding agencies lose patience will be unwilling to try too many layers, and will thus tend to (literally or metaphorically) hard-wire the parts they are less interested in. I would expect that a category-formation experiment would have to involve more layers than (and include) a category-assignment experiment. -David West dhw@iti.org