Powered by Indico
v3.2.8
Flatiron Institute community members are cordially invited to a CCN Seminar with Dimitra Maoutsa (Postdoc, TU Munich). Her talk title and abstract are below.
Title: Learning latent low-dimensional stochastic dynamics in neural population activity: an optimal control approach
Abstract:
A fundamental challenge in systems neuroscience is understanding how cognitive and behaviourally relevant latent processes are reflected in neuronal population responses. While latent low-dimensional deterministic mechanisms have been instrumental in explaining collective neural activity, they are often unfit to describe the inherent randomness intrinsic to cognitive processes such as decision-making. In my talk, I will present a method for identifying latent stochastic dynamical structure in neural population responses based on stochastic control. To that end, I will first introduce the necessary components for building this framework: i) an interacting particle system that allows for efficient sampling of marginal probability densities of stochastic systems based solely on deterministic dynamics, and ii) a non-iterative stochastic control method that employs the deterministic interacting particle dynamics to compute optimal controls. I will demonstrate how the constraints of the stochastic control framework naturally translate to a likelihood function used in statistical inference problems. I will apply this approach for inference of a dynamical system with direct state observations and geometric constraints, and latent dynamical systems with indirect observations mediated through neural population responses.