Diffusion Bridge Variational Inference for Deep Gaussian Processes
Published in The Fourteenth International Conference on Learning Representations (ICLR 2026), 2026
We extend score-based variational inference for Deep Gaussian Processes by inserting a Doob h-transform into the forward noising SDE so the diffusion interpolates between an amortized, data-anchored initial distribution and a fixed terminal noise. The bridge marginal is Gaussian in closed form, so a conditional score network can be trained against it via denoising score matching, and posterior samples are produced by integrating the reverse-time bridge SDE.
Recommended citation: Xu, J., Zeng, D., Zhao, Q. and Paisley, J. (2026). Diffusion Bridge Variational Inference for Deep Gaussian Processes. ICLR 2026.
Download Paper