Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Surprisal as a continuously travelling wave across cortex
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Pierre Guilleminot1,2, Benjamin Morillon1,2; 1INSERM, 2Aix-Marseille Université
The brain processes language by continuously integrating contextual expectations, yet the temporal architecture of this process remains debated. Classical electrophysiology describes predictive processing through discrete, stereotyped evoked response potential (ERP) components (MMN, N400…), implying a staged, modular organization. Whether these components reflect genuine processing stages or artifacts of spatial and temporal averaging is an open question. Here, we leverage the excellent signal-to-noise ratio and spatio-temporal resolution of intracortical recordings to characterize how and when the brain encodes predictive information during naturalistic speech comprehension. We analyzed intracortical recordings from 33 subjects with epilepsy, listening to 10 minutes of continuous speech. Word-level surprisal was extracted using a large language model. Neural responses to surprisal were characterized in two electrophysiological signals, high-frequency activity (HFA, 70–150 Hz) and broadband power, using mutual information. For each channel showing a significant response, we extracted the peak response latency to map how the surprisal signal propagates across the cortex. Individual channels exhibit highly variable peak response latencies, largely independent of word duration or degree of surprisal. When aggregated across channels, the resulting latency distribution is broadly consistent with classical ERP timings, yet this correspondence emerges from the superposition of heterogeneous, spread-out single-channel dynamics rather than from discrete processing stages commonly observed on multiple channels. HFA and broadband power both carry surprisal information, but are not systematically correlated with one another within the same channel. Spatially, broadband responses to surprisal are widely distributed, while HFA responses are more focal, concentrated in auditory areas. By clustering channels based on the latency of their response to surprisal in the broadband signal, we revealed through a functional connectivity analysis that surprisal is being carried through recurrent loops across brain regions. Taken together, these results suggest that what classical averaging presents as discrete processing stages is better described as a continuum of latencies distributed across a heterogeneous cortical landscape. Moreover, while surprisal is being encoded by both the HFA and broadband signals, these neural responses differ greatly, with the broadband signal hinting at recurrent processing loops across brain regions, while the HFA reflects the actual computation of an error signal. Overall this work offers a more granular account of how the brain tracks linguistic expectations, with implications for models of cortical language organization and for the interpretation of macroscale electrophysiological signatures.
Topic Areas: Speech Perception,