Malleable Kernel Interpolation for Scalable Structured Gaussian Process

Conference Paper (2024)
Author(s)

Hanyuan Ban (TU Delft - Signal Processing Systems, TU Delft - Electrical Engineering, Mathematics and Computer Science)

E.H.J. Riemens (TU Delft - Signal Processing Systems, TU Delft - Electrical Engineering, Mathematics and Computer Science)

R.T. Rajan (TU Delft - Signal Processing Systems)

Research Group
Signal Processing Systems
DOI related publication
https://doi.org/10.23919/EUSIPCO63174.2024.10715101
More Info
expand_more
Publication Year
2024
Language
English
Research Group
Signal Processing Systems
Bibliographical Note
Green Open Access added to TU Delft Institutional Repository ‘You share, we take care!’ – Taverne project https://www.openaccess.nl/en/you-share-we-take-care Otherwise as indicated in the copyright section: the publisher is the copyright holder of this work and the author uses the Dutch legislation to make this work public. @en
Pages (from-to)
997-1001
ISBN (electronic)
9789464593617
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract

Gaussian process regression (GPR), is a powerful non-parametric approach for data modeling, which has garnered considerable interest in the past decade, however its widespread application is impeded by the significant computational burden for larger datasets. The computational complexity for both inference and hyperparameter learning in GPs lead to O(N3) for N training points. The current state-of-the-art approximations, such as structured kernel interpolation (SKI)-based methods e.g., Kernel Interpolation for Scalable Structured Gaussian Process (KISSGP), have emerged to mitigate this challenge by providing a scalable inducing point alternatives. However, the choice of the optimal number of grid points, which influences the accuracy and efficiency of the model, typically remains fixed and is chosen arbitrarily. In this work, we introduce a novel approximation framework, Malleable KISSGP (MKISSGP), which dynamically adjusts grid points using a new hyperparameter of the model called density, which adapts to the changes in the kernel hyperparameters in each training iteration. In comparison with the state-of-the-art KISSGP and irrespective of changes in hyperparameters, our proposed MKISSGP algorithm exhibits consistent error levels in the reconstruction of the kernel matrix, and offers reduced computational complexity. We present extensive simulations to validate the improved performance of the proposed MKISSGP, and give directions for future research.

Files

Malleable_Kernel_Interpolation... (pdf)
(pdf | 0 Mb)
- Embargo expired in 28-04-2025
License info not available