A small number of common factors often explain most of the interdependence among simultaneously recorded neurons, a signature of underlying low-dimensional dynamics. We posit that simple neural coding and computation manifest as low-dimensional nonlinear dynamics implemented redundantly within a large population of neurons. Recovering the latent dynamics from observations can offer a deeper understanding of neural computation. We improve upon previously-proposed methods for recovering latent dynamics, which assume either an inappropriate observation model or linear dynamics. We propose a practical and efficient inference method for a generative model with explicit point process observations and an assumption of smooth nonlinear dynamics. We validate our method on both simulated data and population recording from primary visual cortex.
View on arXiv