• Modeling Neural-Behavioral Dynamics in C. elegans with RNN-Based State Space and Video Generative Models

Modern recording techniques have enabled large-scale measurements of neural activity alongside the exhibited behaviors across various model organisms. The dynamics of neural activity shed light on how organisms process sensory information and generate motor behaviors. In this study, we investigated these dynamics using optical recordings of neural activity in the nematode C. elegans.

We developed state space models using Recurrent Neural Networks (RNNs) to analyze neural activity data and predict the worm’s behaviors. Another core component of the project was the creation of a video generative model (VGM) capable of predicting future frames of the worm movement based on an initial sequence. By integrating visual and temporal data, the model forecasted the worm’s behavior over time, offering a dynamic perspective on how the behaviors evolve.

The embeddings from the behavior predictor and the latent spaces of the video generative model were subsequently analyzed to explore the correlations between neural activity and behavior. This analysis revealed the underlying structure of neural data and its connection to the observed behaviors, providing critical insights into the fundamental principles of neural processing and behavior.