When do memory limitations lead to regularization? An experimental and computational investigation

Abstract

Elsevier believes that individual authors should be able to distribute their AAMs for their personal voluntary needs and interests, e.g. posting to their websites or their institution's repository, e-mailing to colleagues. However, our policies differ regarding the systematic aggregation or distribution of AAMs to ensure the sustainability of the journals to which AAMs are submitted. Therefore, deposit in, or posting to, subject-oriented or centralized repositories (such as PubMed Central), or institutional repositories with systematic posting mandates is permitted only under specific agreements between Elsevier and the repository, agency or institution, and only consistent with the publisher's policies concerning such repositories. Voluntary posting of AAMs in the arXiv subject repository is permitted. nd April 2013When do memory limitations lead to regularization? An experimental and computational investigation Amy PerforsSchool of Psychology, University of Adelaide, Adelaide, SA 5005, Australia AbstractThe Less is More hypothesis suggests that one reason adults and children differ in their ability to learn language is that they also differ in other cognitive capacities. According to one version of this hypothesis, children's relatively poor memory may make them more likely to regularize inconsistent input Newport, 2005, 2009). This paper reports the result of an experimental and computational investigation of one aspect of this version of the hypothesis.A series of seven experiments in which adults were placed under a high cognitive load during a language-learning task reveal that in adults, increased load during learning (as opposed to retrieval) does not result in increased regularization. A computational model offers a possible explanation for these results. It demonstrates that, unless memory limitations distort the data in a particular way, regularization should occur only in the presence of both memory limitations and a prior bias for regularization. Taken together, these findings suggest that the difference in regularization between adults and children may not be solely attributable to differences in memory limitations during learning.

Language learning, language use and the evolution of linguistic variation

“…A common hypothesis is that children's bias toward regularization might be due to their limited memory [4–6]. However, these accounts are hard to reconcile with other research indicating that limitations of this type do not necessarily lead to more regularization [7,8]; while it is possible that memory limitations may play a role, it seems unlikely that they are the main driving force behind this behaviour. Consistent with this, there is evidence that learners bring domain-specific biases to the language learning task: experimental paradigms which compare this tendency to regularize in closely matched linguistic and non-linguistic tasks indicate stronger biases for regularity in language [9,10], suggesting that learners may expect language or communicative conventions more generally not to exhibit unpredictable variation (a point we return to in §4).…”

More Is More in Language Learning: Reconsidering the Less‐Is‐More Hypothesis

“…Hudson Kam and Chang () showed that the amount of regularization decreased when learners were given flashcards with words printed on them to reduce demands on lexical retrieval. Yet, conversely, manipulations designed to increase cognitive load did not increase regularization: When Perfors () tested adults under six different load conditions as they learned an artificial language with complex patterns of inconsistency (modeled after Hudson Kam & Newport, ), load groups were no more likely to regularize determiner use than a no‐load control group. Indeed, the data showed a nonsignificant trend in the opposite direction, with less regularization under concurrent load, counter to predictions of the Less‐Is‐More hypothesis.…”

“…In such cases, the network may show effects of catastrophic interference where patterns of connectivity among hidden units become entrenched during learning, hindering subsequent learning that requires establishing a different pattern of connectivity. Further doubt about the benefits of capacity limitations comes from statistical models: Perfors () explored effects of capacity limitations by simulating learning as a process of Bayesian inference. She demonstrated that when capacity is diminished, regularization takes place only if an a priori bias toward regularization is present, which suggests that regularization is not a result of capacity limitations per se.…”

“…Such findings were taken as support for the view that children play a crucial role in systematizing and regularizing inconsistent language input as it is transmitted over generations. However, Perfors (2012aPerfors ( , 2016 showed that when adults experienced less pressure to be correct, they regularized the input just as much as children-leaving open the possibility that the observed differences between adults and children in regularization may be due to differences in pragmatic and social assumptions about the goals of the task and the nature of the encountered variation, rather than cognitive limitations.…”

Adults are more efficient in creating and transmitting novel signalling systems than children

“…Based on this view, structure should emerge more readily in children than in adults during iterated language learning. However, it is not clear whether cognitive immaturity per se aids structure-inducing innovations as attempts to demonstrate experimentally that cognitive limitations lead to superior decomposition (Cochran, McDonald & Parault, 1999) or regularisation of input (Perfors, 2012a) have proved unsuccessful or are open to alternative interpretations (Perfors, 2012b;Rohde & Plaut, 1999;2003).…”