berg@cs.albany.edu (George Berg) (06/25/91)
Our story so far... >In article <1991Jun24.181310.17421@cs.ubc.ca> leblanc@cs.ubc.ca (David LeBlanc) writes: >> In article <857@odin.albany.edu> berg@odin.albany.edu.UUCP (George Berg) I hate to get any further involved in this debate, but, what the heck it's Summer... Some specific responses: >> As to their inabiltity to represent recursion, your information is out of >>date. Those who are interested in an update on the debate should look at Tim >>van Gelder's "Compositionality: A Connectionist Variation on a Classical >>Theme" in the journal Cognitive Science V. 14, N. 3 (1990). van Gelder >>discusses no fewer than three strategies (Pollack, Smolensky, and Hinton) >>researchers are using to represent structured information. > >While it is true that, given enough tweaking and pulling, it may be possible >to get connectionist networks to do pretty well anything, the fact that >there is a far simpler paradigm for approaching these problems undermines >any performance argument for a connectionist approach. Tweaking and pulling? Hardly. As the paper shows, these architectures are not hasty kludges. To the contrary, they represent interesting alternatives, which connstitute pretty strong performance arguments for connectionist models of human language *performance* with its clearly bounded capabilities. (I apologize for reprinting my entire original paragraph, but I still recommend the van Gelder paper to those people who want to see one of the better connectionist rebuttals to the Fodor and Pylyshyn (Cognition, 1988) critiques.) >The problem with connectionism seems to be that every charlaten and academic >slight-of-hand artist is jumping on the bandwagon. > [accounts of stupid things misinformed people say about connectionism deleted] >These are the people who give >connectionism its bad name. Nobody is more infuriated by these people than those of us trying to do serious research. Their stupid claims make it harder for us to convince others of our (hopefully) more sober ones. The same problem holds true in AI overall, where every few years some moron comes along with a book saying that "in five years we'll have <fill in your favorite ludicrous claim>". However, this is beside the point when it comes to connectionism as a valid AI or cognitive model. I defend my claims about connectionism and those of others whose judgement I find sound, not those of every person who has heard a connectionist talk. >You seem to be advocating a hybrid approach, a stand I am also in favour of. Just to set the record straight, it is my hope that connectionist models, and other, more neurally plausible models which might succeed them, do form the basis for our comprehensive understanding of human knowledge and use of language. One of the bases for my work is that current work in theoretical linguistics (especially the principles and constraints used in GB) provides helpful guidelines as to how connectionist models of language might be organized (without, I might add, relegating them to being "mere implementations"). So, I like both approaches. Some general comments with respect to why connectionist models are worth examining: 1. The early (c. 1985) connectionist language models were good work. As a matter of fact, one of them, the Rumelhart and McClelland past tense learning model demonstrated something important - that complex linguistic phenomena do not need to be represented as explicit rules. Sure, there were those who admitted the abstract possibility, but this brought the point home in a concrete manner. Note that I am not defending their claims of psychological validity here. That's a different issue entirely. A whole segment of the literature has been devoted to batting that one back and forth. 2. Strangling an idea in its crib. Although you can argue that connectionism has done nothing to prove its worth (I claim I disagree with) I would argue that you are stacking the deck. There has been an entire generation of research in modern linguistics and cognitive psychology. In its current incarnation, connectionist research has only blossomed since about 1985 (or so). In fairness, you should give it a few years. And as I stated in my previous message, it's not as if the field has remained static. 3. As Stevan Harnad mentions in his review of the Fodor/Pylyshyn,Pinker/Prince,Lachter/Bever critiques of connectionism (I don't remember where it appeared - either AIj or Connection Science, last year I'm away from my library at the moment), neither the connectionists, *nor* the advocates of physical symbol systems has demonstrated that they have a theory which can produce models even approaching a comprehensive grasp of human cognition. In a sense we are beating each other over the head with clubs made of our hopes and little else ("my theory has the potential for working, but yours doesn't"). I'll grant that at the moment, the symbolists (for lack of a better term) are ahead on points. But there's a long way to go. 4. If connectionism is a dead-end it will go away. The journals devoted to it will dry up. Other journals will refuse to publish connectionist material. Granting agencies will not support connectionist proposals. Connectionists will not find jobs or get tenured or promoted. If you really believe connectionism is a dead end, relax, sit back, wait five years and then you can send a message to sci.lang saying that you were right. :-) George P.S. If anyone else wants to help with the pro-connectionist side in this debate, they're more than welcome (hint, hint). Frankly, I'm not too comfortable being point man here, and I am sure others can defend connectionism much more persuasively and eloquently than can I. ------------------------------------------------------------------------------- | George Berg | Computer Science Dept. | If you want wit in 15 words | | berg@cs.albany.edu | SUNY at Albany, LI 67A | or less, go check Bartlett's | | (518) 442 4267 | Albany, NY 12222 USA | quotations -- I'm busy. | -------------------------------------------------------------------------------