Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Contributed Talk Session: Thursday, August 14, 11:00 am – 12:00 pm, Room C1.03
Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Decomposition of uncertainty into dispersion and strength during speech processing

Pierre Guilleminot¹, Benjamin Morillon; ¹Université d'Aix-Marseille

Presenter: Pierre Guilleminot

Speech comprehension relies on contextual predictions to minimize error between expected and incoming auditory input. This mechanism hinges on linguistic uncertainty, typically quantified using Shannon entropy, which captures the overall unpredictability of a given word based on prior context. However, Shannon entropy is only one member of the broader Rényi entropy family, which enables a distinction between uncertainty due to the strength of dominant predictions and that due to the dispersion across alternatives. Here, we investigated which entropy-based measure of uncertainty best reflects neural processing during speech listening. Using intracortical recordings from subjects listening to an audiobook, we computed the mutual information between neural responses and different uncertainty measures derived from a large language model. Our results reveal that, rather than Shannon entropy, the brain separately processes the strength and dispersion of predictions in distinct neural populations. This suggests a multidimensional representation of uncertainty. More generally, these findings highlight the need for a more refined definition of uncertainty in cognitive neuroscience.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF