Schedule of Events | Symposia

Sketchpad Series

Learning from our mistakes? The role of prediction errors in statistical learning: an eye-tracking study

Poster Session B - Sunday, March 30, 2025, 8:00 – 10:00 am EDT, Back Bay Ballroom/Republic Ballroom

Flóra Hann^1,2,3 (hannflora@gmail.com), Cintia Anna Nagy², Dezső Németh^4,5,6, Orsolya Pesthy^2,7; ¹Doctoral School of Psychology, ELTE Eötvös Loránd University, Budapest, Hungary, ²Institute of Psychology, ELTE Eötvös Loránd University, Budapest, Hungary, ³Institute of Experimental Medicine, HUN-REN Research Centre for Natural Sciences, Budapest, Hungary, ⁴Centre de Recherche en Neurosciences de Lyon CRNL U1028 UMR5292, INSERM, CNRS, Université Claude Bernard Lyon 1, Bron, France, ⁵Department of Education and Psychology, Faculty of Social Sciences, University of Atlántico Medio, Las Palmas de Gran Canaria, Spain, ⁶BML-NAP Research Group, Institute of Psychology, Eötvös Loránd University and Institute of Cognitive Neuroscience and Psychology, HUN-REN Research Centre for Natural Sciences, Budapest, Hungary, ⁷Institute of Cognitive Neuroscience and Psychology, HUN-REN Research Centre for Natural Sciences, Budapest, Hungary

The framework of Bayesian inference is often applied to statistical learning. It suggests that learning is driven by mismatches between predictions and outcomes (i.e., prediction errors). These errors induce the updating of representations (priors) about the underlying regularities of the environment. However, our environment is not entirely deterministic, thus, some errors reflect the presence of noise (learning-dependent errors) rather than an inaccurate representation of regularities (not-learning-dependent errors). Therefore, not all errors should be weighted equally. The role of these errors in statistical learning remains unclear, therefore we tested how their different types induce the updating of representations, and how this drives statistical learning. We used the gaze-contingent eye-tracking version of a statistical learning task, where we assessed predictions by registering anticipatory eye-movements. To distinguish between learning-dependent and not-learning-dependent predictive errors, we chose a task with a probabilistic structure. We expected the likelihood of learning-dependent errors to increase as time progressed and participants learned the regularity. We also expected not-learning-dependent errors to induce the updating of representations more than learning-dependent errors, as the former reflect inaccurate representations. Preliminary analyses of anticipatory eye-movements suggest that learning-dependent errors were the most likely compared to other anticipation types during the task. Contrary to expectations, there was no significant difference between error types in inducing updating. Thus, mechanisms behind updating need further investigation. In summary, analyzing the likelihood and updating of different types of prediction errors provides a promising tool to uncover the mechanisms underlying statistical learning.

Topic Area: LONG-TERM MEMORY: Skill Learning