Scalable Log Determinants for Gaussian Process Kernel Learning
Authors: Kun Dong, David Eriksson, Hannes Nickisch, David Bindel, Andrew G. Wilson
NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We illustrate the performance of our approach on several large, multi-dimensional datasets, including a consequential crime prediction problem, and a precipitation problem with n = 528, 474 training points. We consider a variety of kernels, including deep kernels [24], diagonal corrections, and both Gaussian and non-Gaussian likelihoods. |
| Researcher Affiliation | Collaboration | 1 Cornell University, 2 Phillips Research Hamburg |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We have released code and tutorials as an extension to the GPML library [25] at https: //github.com/kd383/GPML_SLD. A Python implementation of our approach is also available through the GPy Torch library: https://github.com/jrg365/gpytorch. |
| Open Datasets | Yes | This experiment involves precipitation data from the year of 2010 collected from around 5500 weather stations in the US1. The hourly precipitation data is preprocessed into daily data if full information of the day is available. The dataset has 628, 474 entries in terms of precipitation per day given the date, longitude and latitude. We randomly select 100, 000 data points as test points and use the remaining points for training. 1https://catalog.data.gov/dataset/u-s-hourly-precipitation-data |
| Dataset Splits | No | The paper describes training and testing splits for datasets (e.g., 'We randomly select 100, 000 data points as test points and use the remaining points for training.' for precipitation data) but does not mention a separate validation split or cross-validation. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions 'GPML library [25]' and 'GPy Torch library' but does not specify version numbers for these or any other software components. |
| Experiment Setup | Yes | We use 5 probe vectors and 25 iterations for Lanczos, both when building the surrogate and for hyperparameter learning with Lanczos. We also use 5 probe vectors for Chebyshev and 100 moments. |