Reproduce “Multi-resolution Multi-task Gaussian Processes” on OpenResearch

About this paper

Oliver Hamelijnck, Theodoros Damoulas, Kangrui Wang, Mark Girolami

Researchers at The Alan Turing Institute, University of Warwick, and University of Cambridge developed a Multi-resolution Multi-task Gaussian Process (MRGP) framework, encompassing both shallow (MR-GPRN) and deep (MR-DGP) models, for integrating heterogeneous spatio-temporal data from multiple observation processes. The framework accounts for dependencies and biases, achieving improved predictive accuracy and uncertainty quantification in synthetic and real-world air pollution prediction tasks.

Key insights

• Information-theoretic corrections applied to composite likelihoods can effectively mitigate posterior contraction, providing more accurate uncertainty quantification when modeling dependent observation processes.
• A layered, tree-structured Deep Gaussian Process architecture offers a robust mechanism for learning and correcting systematic biases between heterogeneous data sources with varying resolutions or fidelities.
• Approximating integrals over sampling areas via discretization allows for flexible integration of multi-resolution observations into GP models without requiring specific kernel types or single-resolution assumptions.

Summary by alphaXiv.