Variational Fourier Features for Gaussian Processes

Authors: James Hensman, Nicolas Durrande, Arno Solin

JMLR 2017 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Section 6 is dedicated to a set of illustrative toy examples and a number of empirical experiments, where practical aspects of Variational Fourier Feature inference are demonstrated. ... Table 2 shows the predictive mean squared errors (MSEs) and the negative log predictive densities (NLPDs) with one standard deviation on the airline arrival delays experiment.
Researcher Affiliation Collaboration James Hensman EMAIL Nicolas Durrande EMAIL PROWLER.io 66-68 Hills Road Cambridge, CB2 1LA, UK Arno Solin EMAIL Aalto University FI-00076 AALTO Espoo, Finland
Pseudocode No The paper describes methodologies and mathematical derivations but does not include any explicitly labeled pseudocode or algorithm blocks with structured steps in a code-like format.
Open Source Code Yes All of the methods in this paper and all of the code to replicate the experiments in Section 6 are available at http://github.com/jameshensman/VFF.
Open Datasets Yes The US flight delay prediction example (see Hensman et al., 2013, for the original example) has reached the status of a standard test data set in Gaussian process regression... We repeat an experiment from that work on solar irradiance data (Lean, 2004)
Dataset Splits Yes Predictions are made for several subset sizes of the data... In each case, two thirds of the data is used for training and one third for testing.
Hardware Specification Yes Running the VFF experiment with all 5.93 million data using our Python implementation took 265 ± 6 seconds (626 ± 11 s CPU time) on a two-core Mac Book Pro laptop (with all calculation done on the CPU).
Software Dependencies No We made use of Tensor Flow (Abadi et al., 2015) and GPflow (Matthews et al., 2017) to construct model classes in Python. Specific version numbers for these software dependencies are not provided.
Experiment Setup Yes We started by fitting a full GP model... maximising the likelihood with respect to the kernel s variance parameter, lengthscale parameter and the noise variance. ... We drew data from a Gaussian process with a Mat ern-3 2 kernel (σ2 = 1, ℓ= 0.2) and added Gaussian noise (σ2 n = 0.05). ... For the Variational Fourier Feature method (VFF), we used M = 30 frequencies per input dimension. ... In the VFF method, the inputs were normalized to [0, 1] and the domain size was set to [−2, 3]...