GPyTorch: Blackbox Matrix-Matrix Gaussian Process Inference with GPU Acceleration
Authors: Jacob Gardner, Geoff Pleiss, Kilian Q. Weinberger, David Bindel, Andrew G. Wilson
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments we show that BBMM effectively uses GPU hardware to dramatically accelerate both exact GP inference and scalable approximations. Additionally, we provide GPy Torch, a software platform for scalable GP inference via BBMM, built on Py Torch. |
| Researcher Affiliation | Academia | Jacob R. Gardner , Geoff Pleiss , David Bindel, Kilian Q. Weinberger, Andrew Gordon Wilson Cornell University {jrg365,kqw4,andrew}@cornell.edu, {geoff,bindel}@cs.cornell.edu |
| Pseudocode | No | The paper describes algorithms such as 'Modified CG' and 'Pivoted Cholesky Algorithm' in text, but it does not include a formal pseudocode or algorithm block. |
| Open Source Code | Yes | We introduce GPy Torch, a new software platform using BBMM inference for scalable Gaussian processes, which is built on top of Py Torch: https://gpytorch.ai. |
| Open Datasets | Yes | We test Exact models on five datasets from the UCI dataset repository [2] with up to 3500 training examples (the largest possible before all implementations exhausted GPU memory): Skillcraft, Gas, Airfoil, Autompg, and Wine. [2] A. Asuncion and D. Newman. Uci machine learning repository. https://archive.ics.uci.edu/ ml/, 2007. Last accessed: 2018-05-18. |
| Dataset Splits | No | The paper mentions using training examples and datasets but does not explicitly provide details about train/validation/test splits, specific percentages, or sample counts for validation. |
| Hardware Specification | Yes | All speed experiments are run on an Intel Xeon E5-2650 CPU and an NVIDIA Titan Xp GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch', 'Adam' optimizer, and 'GPFlow' but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | Experiment details. All methods use the same optimizer (Adam) with identical hyperparameters. In BBMM experiments we use rank k = 5 pivoted Cholesky preconditioners unless otherwise stated. We use a maximum of p = 20 iterations of CG for each solve, and we use t = 10 probe vectors filled with Rademacher random variables to estimate the log determinant and trace terms. SGPR models use 300 inducing points. SKI models use 10,000 inducing points and the deep kernels described in [52]. |