Measuring and predicting structure processing in sounds and language

Structure detection and manipulation are essential to our cognitive abilities. We analyze sounds based on their structure and communicate through sentences structured according to our language syntax. To understand how structures are processed, researchers perform neural recording or turn to artificial models, including large language models (LLM). Our two speakers will successively focus on these two aspects to deliver a precise and broad overview of structure processing, in the brain and as modeled by LLM.

10h30 - Maria Chait (UCL):  "How the brain tracks structure in rapidly unfolding sounds"

Professor Chait studies how biological neural networks detect regularity. For that purpose, she has led several experimental investigations measuring the automatic detection of complex auditory structures in Magneto-encephalography (Barascud et al. 2016). Her work additionally questioned how these regularities are involved in auditory scene analysis. Recently, Professor Chait supervised theses concerned with novel variants of these structure detection, as well as tests of these paradigms in animal models including the Ferret, extending her analysis to single-cell recordings. Her presentation will consequently dissect in detail the neuronal processing underlying rapid structure detection.

11h30 - Micha Heilbron (University of Amsterdam): "Language models for cognitive (neuro)science: Two success stories and a warning"

Professor Heilbron studies how artificial neural networks can be used as a model of cognitive processing. He is an expert on predictive coding as well as encoding models (prediction of brain activity from model activity). His work spans across music (Kern et al. Elif 2022) and language (Heilbron et al. PNAS 2022). Recently, he has been questioning if language models and the brain perform similar predictions. Although next-word prediction is the key element of good encoding models, it remains unknown if biological and artificial neural networks implement the same predictive coding algorithm. Professor Heilbron will consequently present two detailed analyses of how good language models are in modeling results from cognitive neuroscience.

Together, Professors Chait and Heilbron have published an important review on predictive coding and their work will complement each other in suggesting how this predictive coding participates in the process of structure detection.

Heilbron M, Chait (2018). M. Great Expectations: Is there Evidence for Predictive Coding in Auditory Cortex? Neuroscience. 389:54-73. doi: 10.1016/j.neuroscience.2017.07.061. PMID: 28782642.