Scalable Gaussian Processes with Grid-Structured Eigenfunctions (GP-GRIEF)

Authors: Trefor Evans, Prasanth Nair

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We benchmark our algorithms on real-world problems with up to twomillion training points and 1033 inducing points. 6. Experimental Studies
Researcher Affiliation Academia 1University of Toronto, Canada. Correspondence to: Trefor W. Evans <trefor.evans@mail.utoronto.ca>, Prasanth B. Nair <pbn@utias.utoronto.ca>.
Pseudocode Yes Algorithm 1 Computes Sp, and logλp (the log of the p largest eigenvalues of KU,U).
Open Source Code Yes Using the authors code4, we report the mean and standard devi- 4https://github.com/treforevans/gp_grief
Open Datasets Yes We next assess performance on real-world regression datasets from the UCI repository.
Dataset Splits Yes we report the mean and standard deviation of the RMSE from 10-fold cross validation5. 590% train, 10% test per fold. We use folds from https://people.orie.cornell.edu/andrew/code
Hardware Specification Yes Also presented is the mean training time per fold on a machine with two E5-2680 v3 processors.
Software Dependencies No The paper mentions software like 'py-mcmc' and 'GPy' in the references but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes Before training the GP-GRIEF models, we initialize the base kernel hyperparameters, θ, by maximizing the marginal likelihood of an exact GP constructed on minpn, 1000q points randomly selected from the dataset. For all studies, we fix sm 10, and we fix p 1000 for GP-GRIEF-I. The training time includes MCMC sampling, which we run for 10000 iterations. We use log-normal priors with {mode, variance} of t1, 100u for w, and tσ2 0, 0.04u for σ2, where σ2 0 is the initialized value. We begin sampling at the prior mode, burning the first 1000 samples and thinning every 50 thereafter.