Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: Hierarchical Inducing Point Gaussian Process for Inter-domain Observations
(Submitted on 28 Feb 2021 (v1), last revised 24 Jun 2021 (this version, v2))
Abstract: We examine the general problem of inter-domain Gaussian Processes (GPs): problems where the GP realization and the noisy observations of that realization lie on different domains. When the mapping between those domains is linear, such as integration or differentiation, inference is still closed form. However, many of the scaling and approximation techniques that our community has developed do not apply to this setting. In this work, we introduce the hierarchical inducing point GP (HIP-GP), a scalable inter-domain GP inference method that enables us to improve the approximation accuracy by increasing the number of inducing points to the millions. HIP-GP, which relies on inducing points with grid structure and a stationary kernel assumption, is suitable for low-dimensional problems. In developing HIP-GP, we introduce (1) a fast whitening strategy, and (2) a novel preconditioner for conjugate gradients which can be helpful in general GP settings. Our code is available at https: //github.com/cunningham-lab/hipgp.
Submission history
From: Luhuan Wu [view email][v1] Sun, 28 Feb 2021 04:20:58 GMT (2677kb,D)
[v2] Thu, 24 Jun 2021 06:02:27 GMT (2670kb,D)
Link back to: arXiv, form interface, contact.