Reproduce “WFR-MFM: One-Step Inference for Dynamic Unbalanced Optimal Transport” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinyu Wang, Ruoyu Wang, Qiangwei Peng, Peijie Zhou, Tiejun Li
WFR-MFM introduces a mean-flow framework for dynamic unbalanced optimal transport, enabling one-step inference and addressing the computational bottleneck of iterative ODE simulations. This method achieves two to three orders of magnitude faster inference on single-cell RNA sequencing data while preserving predictive accuracy, allowing for scalable analysis of mass-varying biological dynamics and large-scale perturbation prediction.
Key insights
- Summarizing continuous-time dynamics through mean velocity and mass-growth fields allows for direct, one-step state predictions, bypassing iterative ODE simulations during inference.
- Derivative identities linking mean-flow variables to instantaneous dynamics enable a simulation-free training paradigm for these mean-flow networks, extending this efficiency benefit to inference for unbalanced dynamics.
- Learning perturbation-dependent mean-flow fields via conditional modeling facilitates generalization to numerous unseen perturbation conditions, addressing scalability challenges in in silico perturbation prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.