Constraints on Plasticity in a Connectionist Model of the English Past Tense

Abstract

Abstract This paper investigates constraints on dissociation and plasticity in a connectionist model undergoing random "lesions" both prior to and during training. When networks were trained only on phonological encodings of stem-suhed pairs similar to English regular verbs (e.g., walk  walked), long-term deficits (i.e., "critical period" effects) were not observed, yet there were substantive short-term effects of injury. When training vocabulary reflected the English-like competition between regular (suffixed) and irregular verbs (e.g., go  went, hit  hit), the acquisition of regular verbs became increasingly susceptible to injury, while the irregulars were learned quickly and were relatively impervious to damage. Patterns of generalization to novel forms conflicts with the assumption that this behavioral dissociation is indicative of selective impairment of the learning and generalization of the past tense rule, while the associative lexical-based mechanism is left intact. Instead, we propose a view of network performance in which the regular-irregular dissociation derives from a general reduction in the ability to find a single-mechanism solution when resolving the competition between two classes of mappings. In light of other models in which "regular" and "irregular" forms compete (e.g., Patterson, Seidenberg, & McClelland, 1989), as well as patterns of performance in normal and disordered English speakers (e.g., Pinker, 1991), two general implications are discussed: (1) critical period effects need not derive from endog-enously determined maturational change, but instead may in part result from learning history in relation to characteristics of the language to be learned (i.e., entrenchment), and (2) selective dissociations can result from general damage in systems that are not modularized in terms of rule-based vs. associative mechanisms.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 142 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.