Inferring Local Cortical State from Neural Time Series

How does the brain compute? We study electrophysiological signals recorded from primate brain. Neural population implements meaningful computation that eventually produces robust behavior through dynamics. Although behavioral tasks are designed to control the internal states, there are inevitable internal variations, some of which manifest as reaction time distribution, error trials, and change of mind trials. Thus, accessing internal dynamical variables is necessary to understand how cortex computes. However our observation of the dynamics through invasive electrophysiology is indirect, partial, and highly noisy. We use probabilistic models to posit shared low­-dimensional latent dynamics to observations to reveal how multiple functional neural populations dynamically interact with each other. We developed a variational Bayesian method to infer the posterior on the underlying dynamics. Our inference algorithm is memory-­efficient and fast: both linear in duration using a low­-rank approximation of the prior covariance matrix.