Distributional learning is error-driven: the role of surprise in the acquisition of phonetic categories

Abstract

<jats:title>Abstract</jats:title><jats:p>Much previous research on distributional learning and phonetic categorization assumes that categories are either faithful reproductions or parametric summaries of experienced frequency distributions, acquired through a Hebbian learning process in which every experience contributes equally to the category representation. We suggest that category representations may instead be formed via error-driven predictive learning. Rather than passively storing tagged category exemplars or updating parametric summaries of token counts, learners actively anticipate upcoming events and update their beliefs in proportion to how surprising/unexpected these events turn out to be. As a result, rare category members exert a disproportionate influence on the category representation. We present evidence for this hypothesis from a distributional learning experiment on acquiring a novel phonetic category, and show that the results are well described by a classic error-driven learning model (Rescorla, R. A. &amp; A. R. Wagner. 1972. A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement. In A. H. Black &amp; W. F. Prokasy (eds.), <jats:italic>Classical conditioning II: Current research and theory</jats:italic>, 64–99. New York, NY: Appleton-Century-Crofts).</jats:p>

LDL-AURIS: A computational model, grounded in error-driven learning, for the comprehension of single spoken words

“…Cues are sublexical form features, and outcomes are values on the axes of a high-dimensional semantic space. Error-driving learning as formalized by Rescorla and Wagner (1972) has proven to be fruitful for understanding both animal learning (Bitterman, 2000;Gluck and Myers, 2001;Rescorla, 1988) and human learning (Ellis, 2006;Nixon, 2020;Olejarczuk et al, 2018;Ramscar et al, 2014;Ramscar and Yarlett, 2007;Ramscar et al, 2010;Siegel and Allan, 1996).…”

Three steps forward for predictability. Consideration of methodological robustness, indexical and prosodic factors, and replication in the laboratory

“…(Experimental approaches which control morphophonological content or simulate specific speaker or listener effects are increasing, e.g. Lam & Watson 2014, Buz et al 2016, Olejarczuk et al 2018, Tomaschek et al 2018 Facilitated by the availability of large collections of conversational speech and by the development of increasingly sophisticated automatic analysis methods, corpus-based work is growing rapidly, with applications to many fields including sound change (e.g. Hay & Foulkes 2016) and forensic speaker comparison (e.g.…”

Modeling Maltese Noun Plural Classes without Morphemes

“…This presupposes statistical learning in which the strength of the association between a form and its label is determined by the frequency with which labels to similar items are found. There is, however, evidence that learning is not (only) associative, but rather error-driven (Ramscar et al, 2010(Ramscar et al, , 2013aNixon, 2020;Olejarczuk et al, 2018). In the next section, we will discuss how an error-driven model classifies the nouns of the Maltese noun plural system without recourse to morphemes.…”

“…Discriminative learning as proposed by Ramscar and colleagues is based on a long tradition in animal learning (Kamin, 1969;Rescorla and Wagner, 1972;Rescorla, 1988) and has also been successfully applied to language use and language learning, especially processing of morphologically complex words (Baayen et al, 2011(Baayen et al, , 2013(Baayen et al, , 2018aNixon, 2020;Olejarczuk et al, 2018;Arnold et al, 2017;Tomaschek et al, 2019). The mathematics of the model are explained in appendix B.…”