Rethinking Eliminative Connectionism

Gary F. Marcus

Research output: Contribution to journalArticle

Abstract

Humans routinely generalize universal relationships to unfamiliar instances. If we are told "if glork then frum," and "glork," we can infer "frum"; any name that serves as the subject of a sentence can appear as the object of a sentence. These universals are pervasive in language and reasoning. One account of how they are generalized holds that humans possess mechanisms that manipulate symbols and variables; an alternative account holds that symbol-manipulation can be eliminated from scientific theories in favor of descriptions couched in terms of networks of interconnected nodes. Can these "eliminative" connectionist models offer a genuine alternative? This article shows that eliminative connectionist models cannot account for how we extend universals to arbitrary items. The argument runs as follows. First, if these models, as currently conceived, were to extend universals to arbitrary instances, they would have to generalize outside the space of training examples. Next, it is shown that the class of eliminative connectionist models that is currently popular cannot learn to extend universals outside the training space. This limitation might be avoided through the use of an architecture that implements symbol manipulation.

Original languageEnglish (US)
Pages (from-to)243-282
Number of pages40
JournalCognitive Psychology
Volume37
Issue number3
StatePublished - Dec 1998

Fingerprint

Neural Networks (Computer)
symbol
manipulation
Names
Language
language

ASJC Scopus subject areas

  • Experimental and Cognitive Psychology
  • Linguistics and Language

Cite this

Marcus, G. F. (1998). Rethinking Eliminative Connectionism. Cognitive Psychology, 37(3), 243-282.

Rethinking Eliminative Connectionism. / Marcus, Gary F.

In: Cognitive Psychology, Vol. 37, No. 3, 12.1998, p. 243-282.

Research output: Contribution to journalArticle

Marcus, GF 1998, 'Rethinking Eliminative Connectionism', Cognitive Psychology, vol. 37, no. 3, pp. 243-282.
Marcus GF. Rethinking Eliminative Connectionism. Cognitive Psychology. 1998 Dec;37(3):243-282.
Marcus, Gary F. / Rethinking Eliminative Connectionism. In: Cognitive Psychology. 1998 ; Vol. 37, No. 3. pp. 243-282.
@article{23ae04cdb2724504926715b24463a58c,
title = "Rethinking Eliminative Connectionism",
abstract = "Humans routinely generalize universal relationships to unfamiliar instances. If we are told {"}if glork then frum,{"} and {"}glork,{"} we can infer {"}frum{"}; any name that serves as the subject of a sentence can appear as the object of a sentence. These universals are pervasive in language and reasoning. One account of how they are generalized holds that humans possess mechanisms that manipulate symbols and variables; an alternative account holds that symbol-manipulation can be eliminated from scientific theories in favor of descriptions couched in terms of networks of interconnected nodes. Can these {"}eliminative{"} connectionist models offer a genuine alternative? This article shows that eliminative connectionist models cannot account for how we extend universals to arbitrary items. The argument runs as follows. First, if these models, as currently conceived, were to extend universals to arbitrary instances, they would have to generalize outside the space of training examples. Next, it is shown that the class of eliminative connectionist models that is currently popular cannot learn to extend universals outside the training space. This limitation might be avoided through the use of an architecture that implements symbol manipulation.",
author = "Marcus, {Gary F.}",
year = "1998",
month = "12",
language = "English (US)",
volume = "37",
pages = "243--282",
journal = "Cognitive Psychology",
issn = "0010-0285",
publisher = "Academic Press Inc.",
number = "3",

}

TY - JOUR

T1 - Rethinking Eliminative Connectionism

AU - Marcus, Gary F.

PY - 1998/12

Y1 - 1998/12

N2 - Humans routinely generalize universal relationships to unfamiliar instances. If we are told "if glork then frum," and "glork," we can infer "frum"; any name that serves as the subject of a sentence can appear as the object of a sentence. These universals are pervasive in language and reasoning. One account of how they are generalized holds that humans possess mechanisms that manipulate symbols and variables; an alternative account holds that symbol-manipulation can be eliminated from scientific theories in favor of descriptions couched in terms of networks of interconnected nodes. Can these "eliminative" connectionist models offer a genuine alternative? This article shows that eliminative connectionist models cannot account for how we extend universals to arbitrary items. The argument runs as follows. First, if these models, as currently conceived, were to extend universals to arbitrary instances, they would have to generalize outside the space of training examples. Next, it is shown that the class of eliminative connectionist models that is currently popular cannot learn to extend universals outside the training space. This limitation might be avoided through the use of an architecture that implements symbol manipulation.

AB - Humans routinely generalize universal relationships to unfamiliar instances. If we are told "if glork then frum," and "glork," we can infer "frum"; any name that serves as the subject of a sentence can appear as the object of a sentence. These universals are pervasive in language and reasoning. One account of how they are generalized holds that humans possess mechanisms that manipulate symbols and variables; an alternative account holds that symbol-manipulation can be eliminated from scientific theories in favor of descriptions couched in terms of networks of interconnected nodes. Can these "eliminative" connectionist models offer a genuine alternative? This article shows that eliminative connectionist models cannot account for how we extend universals to arbitrary items. The argument runs as follows. First, if these models, as currently conceived, were to extend universals to arbitrary instances, they would have to generalize outside the space of training examples. Next, it is shown that the class of eliminative connectionist models that is currently popular cannot learn to extend universals outside the training space. This limitation might be avoided through the use of an architecture that implements symbol manipulation.

UR - http://www.scopus.com/inward/record.url?scp=0032247463&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0032247463&partnerID=8YFLogxK

M3 - Article

C2 - 9892549

AN - SCOPUS:0032247463

VL - 37

SP - 243

EP - 282

JO - Cognitive Psychology

JF - Cognitive Psychology

SN - 0010-0285

IS - 3

ER -