Fully Bayesian Differential Gaussian Processes through Stochastic Differential Equations

FB-DiffGP couples two stochastic differential equations: a latent-state SDE whose drift/diffusion are the posterior mean/std of a sparse GP layer, and a length-scale SDE whose neural drift evolves the GP’s ARD kernel parameters over diffusion time. This realises a fully Bayesian treatment of the kernel hyperparameters and produces calibrated uncertainty without sacrificing sparse-GP scalability.

Recommended citation: Xu, J., Lin, Z., Chen, M., Yang, J., Zeng, D. and Paisley, J. (2025). Fully Bayesian differential Gaussian processes through stochastic differential equations. Knowledge-Based Systems, 314, p.113187.
Download Paper