CSE Community Seminar | February 28, 2025

Abstract

A central goal of neuroscience is to understand the link between behavior, motion, and neural activity. Recent advances in experimental and machine learning techniques to simultaneously record motion and neural activity provide extraordinary insight into how neural circuits create motion and longer-lived behavioral states. Probabilistic and dynamical models have enabled us to quantify behavior and model motion and neural activity, but they have yet to be used in a manner that jointly captures all these variables and provides the accurate prediction of one given another. Here, we show how a Helmholtz decomposition of the behavioral dynamics can be used to infer an accurate model for behavior and motion from limited data, in combination with generic low-dimensional mode representations. We combine this approach with a probabilistic model for neural activity that is designed to be suitable for the limitations of experimental data. We first demonstrate the utility of this approach on a toy dataset and then apply it to experimental data of C. elegans motion and neural recordings to show that our model captures real-world dynamics and enables the prediction and control of motion from neural activity. Our framework is generic, both at the representation level and at the dynamical modeling level, so we expect our method to be applicable to a range of datasets.