Active Learning Framework for Probability Paths on Wasserstein Space
Researchers introduced a framework for active learning on measure-valued trajectories, addressing the challenge of inferring continuous probability paths from sparse snapshots in domains like single-cell biology. The method uses Linearized Optimal Transport (LOT) to map distributional snapshots into a tangent space compatible with Gaussian Process modeling, enabling epistemic uncertainty quantification. This allows strategic selection of optimal measurement times, overcoming the limitations of standard Euclidean metrics on infinite-dimensional Wasserstein space. The work is published on arXiv with identifier 2605.30625.
Key facts
- Framework extends active experimentation to the space of measures.
- Uses Linearized Optimal Transport (LOT) for mapping snapshots.
- Enables Gaussian Process modeling in tangent space.
- Addresses epistemic uncertainty quantification.
- Targets domains like single-cell biology with destructive data acquisition.
- Published on arXiv with ID 2605.30625.
- Motivated by prohibitive sequencing costs.
- Standard Euclidean metrics are ill-defined on Wasserstein space.
Entities
Institutions
- arXiv