Function-Space Distributions over Kernels

Authors: Gregory Benton, Wesley J. Maddox, Jayson Salkey, Julio Albinati, Andrew Gordon Wilson

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the practicality of FKL over a wide range of experiments: (1) recovering known kernels from data (Section 5.1); (2) extrapolation (Section 5.2); (3) multi-dimensional inputs and irregularly spaced data (section 5.3); (4) multi-task precipitation data (Section 5.4); and (5) multidimensional pattern extrapolation (Section 5.5). We compare to the standard RBF and Matérn kernels, as well as spectral mixture kernels [34], and the Bayesian nonparametric spectral estimation (BNSE) of Tobar [30].
Researcher Affiliation Collaboration Gregory W. Benton 1 Wesley J. Maddox 2 Jayson P. Salkey 1 Júlio Albinati 3 Andrew Gordon Wilson1,2 1Courant Institute of Mathematical Sciences, New York University 2Center for Data Science, New York University 3Microsoft
Pseudocode Yes A full description of the method is in Algorithm 1 in Appendix 2. Following the procedure outlined in Section 4 and detailed in Algorithm 2 in the Appendix...
Open Source Code Yes Code is available at https://github.com/wjmaddox/spectralgp .
Open Datasets Yes Airline Passenger Data We next consider the airline passenger dataset [14] consisting of 96 monthly observations of numbers of airline passengers from 1949 to 1961... We use the multi-task version of FKL in Section 3.4 to model precipitation data sourced from the United States Historical Climatology Network [19]. We use the product kernel described in Section 5.3 with both separate and shared latent GPs for regression tasks on UCI datasets.
Dataset Splits Yes We standardize the data to zero mean and unit variance and randomly split the training and test sets, corresponding to 90% and 10% of the full data, respectively. We conduct experiments over 10 random splits and show the average RMSE and standard deviation.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions 'Py Torch' and 'GPy Torch' as software used but does not provide specific version numbers for these libraries or any other software dependencies.
Experiment Setup Yes For FKL experiments, we use g(ω) with a negative quadratic mean function (to induce an RBF-like prior mean in the distribution over kernels), and a Matérn kernel with ν = 3/2 (to capture the typical sharpness of spectral densities). We use the heuristic for frequencies in the trapezoid rule described in Section 3.1. Using J = 10 samples from the posterior over kernels... This objective can be optimized using any procedure; we use the AMSGRAD variant of Adam as implemented in Py Torch [26].