Analytic interdomain memory for efficient online HiPPO-SVGP

Zhaodong Guo, Naoki Kiyohara, Wenlong Chen, Harrison Bo Hua Zhu, Yingzhen Li

Keywords: Gaussian processes, online learning, sparse variational inference, interdomain inducing variables, HiPPO, random Fourier features, spherical Bessel functions, streaming prediction

TL;DR: We replace the ODE-based interdomain feature construction in one-dimensional HiPPO-SVGP with closed-form spherical-Bessel overlaps, reducing dependence on ODE discretization while preserving predictive performance.

Abstract: Online Gaussian processes are attractive for streaming prediction but exact inference is cubic and sparse online updates can forget early observations. Online HiPPO-SVGP addresses this issue using HiPPO-based interdomain inducing variables, but its interdomain kernel construction relies on temporal ODE recursion of Fourier-Legendre states. We study an analytic replacement for this construction in the one-dimensional HiPPO-LegS setting with stationary kernels. The method expresses the required sine/cosine-Legendre overlap integrals in closed form using spherical Bessel functions, allowing kernel quantities to be evaluated directly at the target horizon. Experiments on Solar Irradiance and COVID-19 mortality streams show comparable predictive performance to ODE-based OHSVGP, while runtime scaling confirms reduced dependence on ODE discretization steps.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 92