This paper introduces a framework that uses inverse scattering theory to learn stable latent dynamics for video prediction, addressing the instability issue in existing approaches. The proposed Stable Inverse Scattering Dynamics (SISD) method reformulates video dynamics learning as an inverse problem in a scattering-transformed latent space, ensuring Lyapunov stability of the latent evolution. The neural architecture combines scattering-based encoding with port-Hamiltonian neural networks for long-term stable predictions.
Key findings
Proposes SISD, a novel framework for stable latent dynamics in video prediction.
Reformulates video dynamics learning as an inverse problem in scattering-transformed space.
Introduces a neural architecture integrating scattering encoders and port-Hamiltonian dynamics networks.
Demonstrates SISD's superior long-horizon prediction accuracy and stability guarantees.
Limitations & open questions
Further research needed to expand the framework's applicability to more complex video datasets.