Full-text

URL: http://crl.ucsd.edu/~elman/Papers/cogsci98.ps.gz

Cached: PDF-295K    PS-1530K    PS.gz-346K

SAVE AS an easy-to-recall long filename:

Filename format: author--year--title   PDF-295K    PS-346K    :: About GZip'd PS

Filename format: author--year--title--journal|proceedings|...--pages   PDF-295K    PS-346K

Related links

Web search: Google Web Search   ::   Google Scholar

Within this site: References (11)

Paper at a Glance

Generalization, simple recurrent networks, and the
emergence of structure
Jeff Elman
Department of Cognitive Science
University of California, San Diego
elman@crl.ucsd.edu

Introduction If human behavior were list­like, accounting for human behavior would be simple: Just enumerate the list of possible stereotypies. Alternatively, if behavior were predictable on the basis of abstract, fully­productive, context­insensitive rules, our task would be different but similarly straightforward: just list the underlying rules. The problem is that most human behaviors seem to lie somewhere between these two possi­ bilities. Neither lists nor rules capture the richness of cognitive behaviors, which are populated by what Plaut, McClelland, Seidenberg, and Patterson have called ``quasi­regular'' domains. There are underlying generalities to most of what we do (or think), but the generalities are typically partial and tempered by qualifications. The problem is not only in knowing when a generalization should apply, but when it should not. Language is a domain which is particularly interesting from this perspective, and the overgeneralization of regularities by children who are learning language is often cited as evidence for the rule­like nature of language. However, although much has been made of the productivity of child language learners, it is also true that children are enormously conservative. Overgeneralizations are exceptional, or at least not as common as would would think given the vast literature on overgeneralizations. If there is any rule, it is that children do not make rules as readily as has been supposed. Our problem in some sense is how to have it both ways: How do we account for productivity and generality, while also accounting for limiting effects of context, item­specific information, frequency, category structure, and so on---effects which themselves range from general and pre­ dictable to purely idiosyncratic? To many of us, connectionist models have seemed ...

@inproceedings{elman98generalizationSimple,
  author={J. L. Elman},
  title={Generalization, simple recurrent networks, and the emergence of structure},
  year={1998},
  address={Mahwah, NJ},
  editor={M.A. Gernsbacher and S.J. Derry},
  publisher={Lawrence Erlbaum Associates},
  booktitle={Proceedings of the Twentieth Annual Conference of the Cognitive Science Society},
  url={http://www.isrl.uiuc.edu/~amag/langev/paper/elman98generalizationSimple.html}
}

HOME   ::  Conference List   ::   Conference Paper	Comments to: junwang4 you-know-at gmail.com	Last update: 2/2/08