Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!ucbvax!decwrl!shelby!Portia!kortge From: kortge@Portia.Stanford.EDU (Chris Kortge) Newsgroups: comp.ai.neural-nets Subject: Re: request for philosophic reactions to connectionism Keywords: ART, Grossberg Message-ID: <1775@Portia.Stanford.EDU> Date: 23 Apr 89 17:40:54 GMT References: <370@eurtrx.UUCP| <18496@gatech.edu> <7894@phoenix.Princeton.EDU> <18504@gatech.edu> <7903@phoenix.Princeton.EDU> <1@bucsb.UUCP> Sender: Chris Kortge Reply-To: kortge@Portia.Stanford.EDU (Chris Kortge) Organization: Stanford University Lines: 50 In article <1@bucsb.UUCP> adverb@bucsb.bu.edu (Josh Krieger) writes: >Please stop dismissing ART out of ignorance! > >Grossberg's papers are exceptionally difficult to understand because >each discovery is dependent on the "minimal" biological building >blocks discovered in the past 30 years. It would be impossible >to understand microprocessor intricacies with only basic knowledge >of transistors; so don't expect to understand Grossberg's work >without a little research and time. ART is a beautiful discovery. > Good analogy. Notice that even though I don't understand microprocessor intricacies, I can still program a computer. Grossberg seems very concerned with the physical implementation of his systems, and thus includes a lot of complicated mathematical detail. There's nothing objectively wrong with this, but a lot of it may be unneccessary for describing the essential information-processing properties, and keeps "ignorant" people at a distance. >The advantages: > [5 advantages deleted] Like Grossberg, you neglect to point out the disadvantages. Categorical representations are inflexible--they allow similarity, and thus generalization, based on only one dimension (two patterns are in the same category, or they are not). Also, unsupervised learning can't guarantee its representations will be relevant to survival of the organism. Using a complex external teacher might be considered cheating, but you at *least* need reinforcement somewhere. >In other words, ART will not be replaced. It will be added to. I agree (surprised?). Backprop has a major drawback, that being the interference problem--present learning is highly disruptive of past learning. This is probably just another way of saying that learning is too slow; in any case, everyone agrees it's a problem. On the other hand, probably not many "backprop people" would agree with me that the solution is to add something to ART. They would rather add something to backprop. My view is that ART is what will need to be added to backprop, so either way of viewing it will work. I think this merger would be much easier if Grossberg's writing style were more objective, less detailed, and more info-processing oriented. He has extremely important ideas, which people are not giving enough attention. >One last comment: Backprop and ART are apples and oranges! Unfortunately, you're right. And it's time to do some genetic engineering! Chris Kortge kortge@psych.stanford.edu