Data-driven reconstruction of head direction circuit dynamics
Written on August 29th, 2026 by Heike Stein
The head direction (HD) system consists of neurons that fire selectively as a function of an animal’s head orientation. By integrating vestibular and visual signals, these cells maintain a stable internal sense of direction. Current circuit models of the HD system typically assume fixed functional forms for the dynamics of latent HD variables. Data-driven models, in contrast, recover latent states and dynamic trajectories without consideration of the known dynamical properties of this well-studied system.
In this project, we use Gaussian process dynamical models to recover HD dynamics directly from data while preserving a mechanistic, circuit-grounded interpretation. We want to learn models that fulfill core criteria:
- data-driven: the dynamics are learned, not assumed.
- interpretable: each term maps to a biologically grounded circuit mechanism.
- modular: each input (vestibular, visual, olfactory, memory, …) influences HD dynamics through separate terms in the equations of latent dynamics.
Team members
- Sepehr Saeedpour
-
Bennet Sakelaris (future)
- Desdemona Fricker (collaborating PI, U Paris Cité)
Relevant publications
in progress
