LuCiD Seminar: Investigating learning mechanisms with miniature artificial language learning

We have one additional LuCiD seminar in March! Join us for an in-person talk on Thursday 23rd March 2023 (11am UK time - in-person & via Zoom). Prof Vsevolod Kapatsinski (University of Oregon) will talk about Investigating learning mechanisms with miniature artificial language learning.

Abstract: An unfortunate consequence of the cognitive revolution has been fragmentation of the study of learning into distinct domains such as language acquisition, causal learning, categorization, and conditioning. As a result, language learning data are still seldom brought to bear on debates about the mechanisms underlying learning in general (i.e., learning theory). I hope to contribute to ‘defragmenting’ the field and enabling a bidirectional flow of ideas between learning theory and language learning by explicitly comparing and testing the predictions of alternative domain-general learning mechanisms using language data.

In this talk, I will present recent work comparing alternative error-driven learning mechanisms. These mechanisms all make use of prediction error but differ in how error is used to update beliefs and change behavior. I focus on the listener’s task of predicting meaning from form, where the formal cues to meaning vary from subphonemic differences between speech sounds in fundamental frequency and voice onset time, to sequences of segments (i.e., morphs). I will present some evidence that prediction error has effects above and beyond updating associations of experienced cues. In particular, it leads learners to re-evaluate the associations of absent cues (Dickinson & Burke, 1996; Kruschke, 2006; van Hamme & Wasserman, 1994): we learn something about the meanings of forms we do not hear. Furthermore, as proposed by learned selective attention theory (Kruschke, 1992), changing associations of experienced cues is supplemented by reweighting of perceptual dimensions. That is, language learners reallocate attention across perceptual dimensions like voice onset time and fundamental frequency if doing so would improve prediction accuracy. I discuss how the ‘standard theory’ of associative learning (Rescorla & Wagner, 1972) can be extended to account for these results.

How to join the seminar: This seminar will take place in person at the University of Manchester, but we will provide a blended approach for anyone who is unable to attend in person. As always the seminar is free to attend & booking isn't required, just get in touch to request the zoom link and don't forget to join the seminar mailing list

Where to find us: A112 Samuel Alexander Building. The Samuel Alexander Building is number 67 on the campus map. A112 is located on the first floor of the Classics wing of the Samuel Alexander Building, this is the wing facing the Alan Gilbert Learning Commons. 

Leave a Comment

* Indicates fields are required