Scalable Exact Inference in Multi-Output Gaussian Processes
Authors: Wessel Bruinsma, Eric Perim, William Tebbutt, Scott Hosking, Arno Solin, Richard Turner
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the efficacy of the method on various synthetic and real-world data sets. We test the OILMM in experiments on synthetic and realworld data sets. |
| Researcher Affiliation | Collaboration | 1University of Cambridge 2Invenia Labs 3British Antarctic Survey 4Alan Turing Institute 5Aalto University 6Microsoft Research. |
| Pseudocode | Yes | Simple algorithms to perform inference and learning in the OILMM are presented in App. A. |
| Open Source Code | Yes | A Python implementation and code to reproduce the experiments is available at https://github. com/wesselb/oilmm. A Julia implementation is available at https://github.com/willtebbutt/OILMMs.jl. |
| Open Datasets | Yes | We consider a subset of the extensive rain forest data set credited to Hubbell et al. (2005); Condit (1998); Hubbell et al. (1999)... Approximately 30 years worth of the ERA- Interim reanalysis temperature data4 (Dee et al., 2011) is smoothed in time... We consider daily exchange rates with respect to USD of the top ten international currencies and three precious materials in the year 2007; we exactly follow Requeima et al. (2019) in the data and setup of the experiment. We consider 256 voltage measurements from 7 electrodes placed on a subject s scalp while the subject is shown a certain image; Zhang et al. (1995) describes the data collection process in detail. We use the OILMM to find relationships between 28 climate simulators4 (see Taylor et al., 2012, for background)... |
| Dataset Splits | No | The paper mentions training and testing splits (e.g., "We train the OILMM and IGPs... on the first 250 months of the data and test on the next 100 months."), but does not explicitly specify a separate validation split or how it was handled for reproducibility. |
| Hardware Specification | Yes | It takes roughly three hours3 to perform 105 iterations of MCMC (circa 105 marginal likelihood evaluations and 106 prior samples) with 20000 bins, demonstrating the feasibility of a computationally demanding choice of approximate inference procedure. 33.6 GHz Intel Core i7 processor and 48 GB RAM |
| Software Dependencies | No | The paper mentions implementations in Python and Julia ("A Python implementation... A Julia implementation...") but does not specify version numbers for these languages or any other software libraries or dependencies used (e.g., PyTorch, TensorFlow, specific GP libraries). |
| Experiment Setup | Yes | For the OILMM, we use a range of numbers of latent processes, up to m = p = 247, and let the basis H be given by the eigenvectors of the kernel matrix over the points in space (Matérn 5/2 with a different length scale for latitude and longitude). For the (O)ILMM, we use m = 3 latent processes with Matérn 1/2 kernels and randomly initialise and learn the basis H. For the (O)ILMM, we use m = 3 latent processes with exponentiated quadratic kernels and randomly initialise and learn H. Hs are the first ms = 5 eigenvectors of a 28 28 covariance matrix Ks between the simulators, and Hr are the first mr = 10 eigenvectors of the kernel matrix over the points in space (Matérn 5/2 with a different length scale for latitude and longitude). |