Orthogonally Decoupled Variational Gaussian Processes

Authors: Hugh Salimbeni, Ching-An Cheng, Byron Boots, Marc Deisenroth

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, our algorithm demonstrates significantly faster convergence in multiple experiments. We empirically assess the performance of our algorithm in multiple regression and classification tasks.
Researcher Affiliation Academia Hugh Salimbeni Imperial College London hrs13@ic.ac.uk Ching-An Cheng Georgia Institute of Technology cacheng@gatech.edu Byron Boots Georgia Institute of Technology bboots@gatech.edu Marc Deisenroth Imperial College London mpd37@ic.ac.uk
Pseudocode No The paper describes update rules and mathematical formulations but does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code and datasets are publicly available. https://github.com/hughsalimbeni/orth_decoupled_var_gps https://github.com/hughsalimbeni/bayesian_benchmarks
Open Datasets Yes Our code and datasets are publicly available. https://github.com/hughsalimbeni/bayesian_benchmarks. We set |γ| = |β| = 500 for all bases and conducted experiments on 3droad dataset (N = 434874, D = 3) for regression with a Gaussian likelihood and ringnorm data (N = 7400, D = 21) for classification with a Bernoulli likelihood.
Dataset Splits No The paper does not explicitly provide specific percentages, sample counts, or a detailed methodology for training, validation, or test dataset splits in the main text.
Hardware Specification Yes We used a computer with a Tesla K40 GPU and found that, in wall-clock time, the orthogonally decoupled basis with |γ| = 3500, |β| = 1500 was equivalent to a coupled model with |β| = 2000 (about 0.7 seconds per iteration) in our tensorflow [1] implementation.
Software Dependencies No The paper mentions 'tensorflow [1] implementation' and 'Adam optimizer [16]' but does not specify their version numbers.
Experiment Setup Yes We make generic choices for hyperparameters, inducing point initializations, and data processing, which are detailed in Appendix F. We set |γ| = |β| = 500 for all bases... and the orthogonally decoupled basis with |γ| = 3500, |β| = 1500 was equivalent to a coupled model with |β| = 2000.