Contextual Verbal Information Modulates Low Frequency Neuronal Population Activity During Face Processing

Abstract Perception of faces is shaped by prior expectations1,2. To address how expectations modulate face processing, we recorded intracranial EEG3,4 across 2836 recording sites in 20 participants as they heard a valenced description about a person, saw that person's face, and rated their impression. Using a two-step approach, we decomposed high- (HFA; 70–170 Hz) and low-frequency (LFA; 1–8 Hz) domains into spatiotemporal motifs using consensus NMF5. Next, we constructed encoding models for each motif, on a trial-by-trial basis to identify traces of expectations during face processing using pretrained language model semantic embeddings. HFA motifs were modality-specific and tracked semantic content predominantly during the audio epoch. In contrast, the effect of valence was seen in LFA motifs, and trial-by-trial neural signals during face epoch were predicted by semantic content across distributed sites. This suggests that semantic information carried by a preceding short narrative is reflected in low-frequency neural activity during subsequent face processing. Our findings offer insights into frequency-specific contributions to how the human brain integrates prior expectations with incoming sensory input.