Reproduce “On Approaches to Building Surrogate ODE Models for Diffusion Bridges” on OpenResearch

About this paper

Maria Khilchuk, Vladimir Latypov, Pavel Kleshchev, Alexander Hvatov

This work introduces SINDy Flow Matching (SINDy-FM) and Diffusion Schrödinger Bridge Matching–NeuralODE (DSBM-NeuralODE) as surrogate ODE models for diffusion bridges, aiming to improve computational efficiency and interpretability. SINDy-FM achieved microsecond inference speeds and orders of magnitude fewer parameters, while DSBM-NeuralODE effectively captured complex dynamics.

Key insights

• Surrogate modeling offers a viable approach to achieve both computational efficiency and interpretability in continuous-time diffusion bridge models.
• Sparse symbolic models, like SINDy-FM, can yield competitive generative performance with significantly reduced parameter counts and near-instantaneous inference times, providing direct interpretability of dynamics.
• Neural ODEs can serve as flexible surrogates for complex, non-linear drift functions within the Schrödinger Bridge framework, with pre-training on diffusion trajectories offering a meaningful initialization.

Summary by alphaXiv.