Exploring Behavior-Relevant and Disentangled Neural Dynamics with Generative Diffusion Models (BeNeDiff)[NeurIPS'2024]

Authors: Yule Wang, Chengrui Li, Weihan Li, and Anqi Wu, Georgia Tech, USA.

Overview

BeNeDiff is a ML tool pipeline that addresses a fundamental challenge in computational neuroscience:

How can we uncover interpretable neural representations associated with distinct behaviors of interest from large‑scale neural recordings?

While traditional behavior decoding models quantify behavioral information in neural data, they often don’t reveal interpretable neural dynamics. BeNeDiff resolves this research gap by:

• Learning a label-informed disentangled neural latent subspace.
• Using video diffusion models to synthesize behavior videos that activate individual latent trajectories, enabling interpretation of its underlying neural dynamics.

The repository provides core implementations of this pipeline and associated neural analysis codes.

Poster