English

DeepRV: Accelerating Spatiotemporal Inference with Pre-trained Neural Priors

Machine Learning 2026-04-01 v3 Machine Learning

Abstract

Gaussian Processes (GPs) provide a flexible and statistically principled foundation for modelling spatiotemporal phenomena, but their O(N3)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 O(N2)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.

Keywords

Cite

@article{arxiv.2503.21473,
  title  = {DeepRV: Accelerating Spatiotemporal Inference with Pre-trained Neural Priors},
  author = {Jhonathan Navott and Daniel Jenson and Seth Flaxman and Elizaveta Semenova},
  journal= {arXiv preprint arXiv:2503.21473},
  year   = {2026}
}

Comments

Code to reproduce all experiments is available in the dl4bi codebase: https://github.com/MLGlobalHealth/dl4bi

R2 v1 2026-06-28T22:36:39.930Z