Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Compositional encoding by single neurons in humans during speech comprehension
Poster Session E, Sunday, September 14, 11:00 am - 12:30 pm, Field House
Ziv Williams1, Joon Kim2, Jill Cai3; 1Harvard Medical School
Humans are capable of conveying exceptionally complex and diverse information through language. Understanding the basic cellular building blocks by which we derive information from language or by which neurons process natural speech, however, has remained a challenge. Here, we used a rare opportunity to record from single cells in the frontotemporal cortex over prolonged durations and employed interpretable deep neural networks in combination with modeling and decoding techniques to discover detailed and diverse linguistic representations by neurons during natural speech processing. We find that whereas certain neurons preferentially tracked basic sound- or word-level features that generalize broadly across contexts, others reflected concepts that varied in content and topic from real-world events and activities to abstract ideas such as organization or time. Rather than mirroring the linguistic signals, these cells provided ‘gestalt’ level representations of the information being conveyed and could coarsely predict the composition of the phrases and sentences heard during speech. These cellular representations could also be projected onto a common conceptual manifold that reflected their latent structure and provided broad coverage of possible concepts as derived from natural language. Together, these findings reveal a highly scalable cellular process that could support our ability to comprehend natural speech and that could enable linguistic inputs not only to be encoded at a granular level but to also be mapped onto a coarser conceptual space that reflects their composition.
Topic Areas: Meaning: Lexical Semantics, Computational Approaches