Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Toward Decoding Meaning: Single-Neuron Semantics in Natural Dialogue
Poster Session C, Saturday, September 13, 11:00 am - 12:30 pm, Field House
Ana Chavez1, Melissa Franch1, Lizzie Mickiewicz, Assia Chericoni, Katya Kabotyanski, Raissa Mathura, Sai Chamarathi, Layth Mattar, Eleonora Bartoli, Andrew Watrous, Nicole Provenza, Sarah Heilbronner, Sameer Sheth, Benjamin Hayden; 1Baylor College of Medicine
Current speech neuroprosthetic systems fall short of harnessing the brain’s full capacity for language. Understanding the neural computations that translate thought into words—and do so seamlessly in multi-speaker conversations—is a formidable challenge. This project aims to accurately characterize the neural mechanisms underlying language production and comprehension using biostatistical modeling and network-level approaches. We analyze naturalistic conversations recorded from patients undergoing intracranial monitoring for epilepsy, enabling us to examine single-neuron activity in real-world social interactions. We demonstrate that a subset of neurons significantly encodes semantic features, as shown by their modulation by contextual word embeddings. By fitting Poisson ridge regression models to neural spike data, we identify neurons whose firing rates are reliably predicted by at least one semantic embedding dimension. Crucially, the naturalistic setting provides rich sampling of both spoken and heard words, allowing us to dissociate the neural subspaces involved in production and comprehension. In both hippocampus and anterior cingulate cortex, we observe semi-orthogonal subspaces for these two processes. Model-based tuning curve correlations reveal a positive—but not complete—overlap, suggesting shared yet speaker-sensitive representations. This speaker-sensitivity persists within semantic categories (e.g., places vs. body parts), indicating that the structure of neural encoding varies across linguistic domains. Together, these findings provide evidence for fine-grained, semantically grounded representations at the single-neuron level during conversation. By uncovering these neural computations, we lay the groundwork for neuroprosthetics that decode both language production and comprehension with greater fidelity.
Topic Areas: Language Production, Speech Perception