Sparse Gaussian Process Hyperparameters: Optimize or Integrate?

Authors: Vidhi Lalchand, Wessel Bruinsma, David Burt, Carl Edward Rasmussen

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In the experiments we demonstrate the feasibility of this scheme relative to several benchmarks and assess regression performance on a 1-dimensional illustrative example and a range of other datasets. We compare our approach across methods on 5 standard small to medium-sized UCI benchmark datasets.
Researcher Affiliation Collaboration Vidhi Lalchand Department of Physics University of Cambridge vr308@cam.ac.uk Wessel P. Bruinsma Microsoft Research AI4Science wbruinsma@microsoft.com David R. Burt LIDS Massachusetts Institute of Technology dburt@mit.edu Carl E. Rasmussen Department of Engineering University of Cambridge cer54@cam.ac.uk
Pseudocode Yes Algorithm 1 Fully Bayesian Sparse GPR with HMC
Open Source Code No No explicit statement or link providing access to the source code for the methodology described in this paper was found.
Open Datasets Yes We compare our approach across methods on 5 standard small to medium-sized UCI benchmark datasets. Following common practice, we use a 20% randomly selected held out test-set [Rossi et al., 2021, Havasi et al., 2018] and scale the inputs and outputs to zero mean and unit standard deviation within the training set (we restore the output scaling for evaluation) [Salimbeni and Deisenroth, 2017].
Dataset Splits Yes RMSE (standard error of mean) evaluated on average of 10 splits with 80% of the data reserved for training. [...] Following common practice, we use a 20% randomly selected held out test-set [Rossi et al., 2021, Havasi et al., 2018]
Hardware Specification Yes All experiments were conducted on an Intel Core i7-10700 CPU @ 2.90GHz x 16.
Software Dependencies No The paper mentions software like 'pymc3 pm.NUTS sampler' and 'tfp.mcmc.Hamiltonian Monte Carlo' and 'GPflow' but does not specify their version numbers.
Experiment Setup Yes Following common practice, we use a 20% randomly selected held out test-set... and scale the inputs and outputs to zero mean and unit standard deviation within the training set... For consistency we initialise all the inducing locations (Z) identically across the methods, i.e. by using the same random subset of training data split. We use the pymc3 pm.NUTS sampler for GPR + HMC and SGPR + HMC, and tfp.mcmc.Hamiltonian Monte Carlo for Joint HMC