Simulation-Free Learning of Population Dynamics with Wasserstein Lagrangian Residuals
Double-Stitch learns Wasserstein Lagrangian mechanics without running a solver at every training step.
The paper targets population dynamics reconstructed from unpaired time snapshots, including systems that gradient flows miss because they can be conservative or periodic. Its method penalizes an equation-of-motion residual along a learned population path, derived from a Clebsch variational principle. On synthetic, single-cell, and ocean vortex datasets, it matches or beats the compared gradient-flow and simulation-based WLM methods on most tasks. Training is reported as 4-14x faster than WLM.
ArXiv · AI/CL/LG's note
The paper targets population dynamics reconstructed from unpaired time snapshots, including systems that gradient flows miss because they can be conservative or periodic. Its method penalizes an equation-of-motion residual along a learned population path, derived from a Clebsch variational principle. On synthetic, single-cell, and ocean vortex datasets, it matches or beats the compared gradient-flow and simulation-based WLM methods on most tasks. Training is reported as 4-14x faster than WLM.
ArXiv · AI/CL/LG's note
score 4