DeepRV: Accelerating Spatiotemporal Inference with Pre-trained Neural Priors

Jhonathan Navott, Daniel Jenson, Seth Flaxman, Elizaveta Semenova

TL;DR: Decoder-only deep generative model for pre-trained spatiotemporal GP priors, enabling accurate and accelerated Bayesian inference.

Abstract:

Gaussian Processes (GPs) provide a flexible and statistically principled foundation for modelling spatiotemporal phenomena, but their $\mathcal{O}(N^3)$ scaling makes them intractable for large datasets. Approximate methods such as variational inference (VI), inducing-point (sparse) GPs, low-rank kernel approximations (e.g., Nystrom methods and random Fourier features), and approximations such as INLA improve scalability but typically trade off accuracy, calibration, or modelling flexibility. We introduce DeepRV, a neural-network surrogate that replaces GP prior sampling, while closely matching full GP accuracy at inference including hyperparameter estimates, and reducing computational complexity to $\mathcal{O}(N^2)$, increasing scalability and inference speed. DeepRV serves as a drop-in replacement for GP prior realisations in e.g. MCMC-based probabilistic programming pipelines, preserving full model flexibility. Across simulated benchmarks, non-separable spatiotemporal GPs, and a real-world application to education deprivation in London (n = 4,994 locations), DeepRV achieves the highest fidelity to exact GPs while substantially accelerating inference.
Code is provided in the dl4bi Python package, with all experiments run on a single consumer-grade GPU to ensure accessibility for practitioners.

Code Dataset Promise: No

Code Dataset Url: https://github.com/MLGlobalHealth/dl4bi

Format Confirmation: I agree that I have read and followed the formatting instructions for the camera ready version.

Submission Number: 189