Deep Learning for Estimating Two-Body Interactions in Mixed-Species Collective Motion

(This is a joint seminar with the Information Theory Study Group.)

Collective motion is a fundamental phenomenon observed in various biological systems, characterized by the coordinated movement of individual entities. Such dynamics are especially crucial in understanding cellular behaviors, which can now be observed at an individual level in complex tissue formations involving multiple types of cells, thanks to recent advancements in imaging technology. To harness this rich data and uncover the hidden mechanisms of such dynamics, we developed a deep learning framework that estimates equations of motion from observed trajectories. By integrating graph neural networks with neural differential equations, our framework effectively predicts the two-body interactions as a function of the states of the interacting entities.
In this seminar, I will first introduce the structure and hyperparameters of our framework. Subsequently, I will detail two numerical experiments. The first is a simple toy model that was employed to generate data for testing our framework to refine the hyperparameters. The second explores a more complex scenario mimicking the collective motion of cellular slime molds, highlighting our model's ability to adapt to mixed-species interactions.