Learning smooth dynamics via Kinematic Interpolants

Key points

• Inferring dynamics from real-world experimental data is inherently ambiguous, as many plausible dynamical systems can explain the same positional observations at measurement times.1
• For higher fidelity modeling of applications ranging from single-cell RNA to vortex dynamics and even flocking of birds, scientific observables increasingly also include velocity estimates that serve to resolve this issue.1
• However, incorporating this rich information in a scalable manner, free of simulation or unnecessary approximations, has remained a difficult challenge.1
• Empirically, we demonstrate that across several scientific datasets, KI achieves superior reconstruction of the underlying dynamics compared to existing baselines while requiring only a fraction of the training time.1

Sources (1)

• [1]Learning smooth dynamics via Kinematic Interpolants
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 06:14 PM

  In this paper, we introduce Kinematic Interpolants (KI), a novel framework that extends the seminal methodology of stochastic interpolants to infer nonlinear dynamics across multiple snapshots in phase space, i.e., observables with position and velocity measurements.

  Inferring dynamics from real-world experimental data is inherently ambiguous, as many plausible dynamical systems can explain the same positional observations at measurement times.

Extractive summary: sentences quoted from the sources.