What is Modularity Good For
Year: 2006
Type: article
Abstract: We compare three types of dual-route associative architectures for learning the English past tense problem.Identical computational resources are used in (1) a pre-specified modular architecture, with a rule mechanism and an exception mechanism; (2) an architecture with two mechanisms that demonstrate emergent specialization of function for regular and exception verbs; and (3) a redundant system, where both mechanisms attempt to learn all verbs.Networks in which regular and exception verbs were learned in emergent or redundant systems showed many of the behaviors thought to derive from the operation of a modular system, and were more effective at learning the training set overall.The prespecified modular solution was least efficient, due to a difficulty in resolving the competition between the two modules when they sought to drive the same output in different ways.The results are discussed in the context of modularity theory.
Institution: Birkbeck, University of London
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Medical Research Council
Award: G8661763244
OpenAlex ID: W2767365474
