Neural manifolds for speech processing

Speech comprehension requires the brain to represent and combine phonetic, lexical, and semantic information across a linguistic hierarchy. Yet how neuronal populations implement these representations and compositions remains unclear. In this talk, I will show how the framework of neural manifolds offers a powerful lens for this question, combining intracranial recordings in the human cortex with population-level analyses and computational modeling.

I will show that linguistic features are encoded as distinct trajectories on low-dimensional manifolds, which can be resolved analytically by fitting recurrent neural networks directly to the neural data. I will then discuss how these trajectory-based representations are combined across levels, from the composition of phonetic features into syllables, to the interactions between phonetic and semantic representations. Together, these perspectives point toward a population-level mechanism for the representation and composition of speech processing.

Host: Michael Pereira.