Core Concepts
Discovering latent force fields in interacting dynamical systems using neural fields.
Abstract
The content discusses the discovery of latent force fields in interacting dynamical systems through the use of neural fields. It introduces the concept of entangled equivariance and proposes a novel architecture that disentangles global field effects from local object interactions. The method, termed Aether, combines neural fields with equivariant graph networks to accurately discover underlying fields and forecast future trajectories. Experiments are conducted on various settings, including static and dynamic fields, showcasing the effectiveness of the proposed approach.
-
Abstract
- Focuses on discovering latent force fields in interacting systems.
- Proposes neural fields to infer hidden forces from observed dynamics.
-
Introduction
- Discusses systems evolving under field effects.
- Introduces equivariant graph networks for learning interactions.
-
Method
- Presents Aether method for field discovery.
- Describes entangled equivariance and global-local coordinate frames.
-
Experiments
- Evaluates Aether on various settings like electrostatic, Lorentz force, traffic scenes, and gravitational n-body problems.
-
Ablation Experiments
- Studies significance of discovered field and sequential architecture.
-
Conclusion
- Summarizes the contributions of Aether in discovering global fields effectively.
Stats
We theorize the presence of latent force fields.
Our experiments show accurate discovery of underlying fields in various settings.
Quotes
"We propose an approximately equivariant graph network that extends equivariant graph networks."
"Our experiments show that explicitly modeling fields is mandatory for effective future forecasting."