Probing Next-Word and Long-Distance Prediction Using Encoding Modelling and MEG
Inés Schönmann, Floris P. de Lange, Micha Heilbron, Donders Institute for Brain, Cognition and Behaviour, Radboud University, Netherlands
Session:
Contributed Talks 4 Lecture
Location:
South Schools / East Schools
Presentation Time:
Sun, 27 Aug, 14:00 - 14:15 United Kingdom Time
Abstract:
Encoding approaches relying on Large Language Models (LLMs) have found evidence for next-word prediction in ECoG and fMRI. Here, using non- invasive MEG, we show that contextualised as well as non-contextualised word embeddings can predict brain responses well before word onset. We find that encoding performance is sensitive to predictions, but—contrary to findings from fMRI—we do not observe signatures of predictions of multiple words ahead.
