Parallel MCMC with Generalized Elliptical Slice Sampling
Authors: Robert Nishihara, Iain Murray, Ryan P. Adams
JMLR 2014 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we compare Algorithm 3 with other parallel MCMC algorithms by measuring how quickly the Markov chains mix on a number of different distributions. Second, we compare how the performance of Algorithm 3 scales with the dimension of the target distribution, the number of cores used, and the number of chains used per core. These experiments were run on an EC2 cluster with 5 nodes, each with two eight-core Intel Xeon E5-2670 CPUs. |
| Researcher Affiliation | Academia | Robert Nishihara EMAIL Department of Electrical Engineering and Computer Science University of California Berkeley, CA 94720, USA Iain Murray EMAIL School of Informatics University of Edinburgh Edinburgh EH8 9AB, UK Ryan P. Adams EMAIL School of Engineering and Applied Sciences Harvard University Cambridge, MA 02138, USA |
| Pseudocode | Yes | Algorithm 1 Elliptical Slice Sampling Update Algorithm 2 Generalized Elliptical Slice Sampling Update Algorithm 3 Building the Approximation Using Parallelism Algorithm 4 Computing the maximum likelihood multivariate t parameters |
| Open Source Code | No | The paper mentions implementing algorithms in Python and using IPython, and cites the Stan library as related work, but does not provide specific access information (link, explicit statement) to the code for the methodology described in this paper. |
| Open Datasets | Yes | Breast Cancer: The posterior density of a linear logistic regression model for a binary classification problem with thirty explanatory variables (thirty-one dimensions) using the Breast Cancer Wisconsin data set (Street et al., 1993). German Credit: The posterior density of a linear logistic regression model for a binary classification problem with twenty-four explanatory variables (twenty-five dimensions) from the UCI repository (Frank and Asuncion, 2010). Ionosphere: The posterior density on covariance hyperparameters for Gaussian process regression applied to the Ionosphere data set (Sigillito et al., 1989). |
| Dataset Splits | No | The paper describes various datasets used for experiments but does not provide specific information about how these datasets were split into training, test, or validation sets for reproducibility. For example, for the real-world datasets like Breast Cancer and German Credit, it mentions using them for posterior density estimation but not explicit data splits. |
| Hardware Specification | Yes | These experiments were run on an EC2 cluster with 5 nodes, each with two eight-core Intel Xeon E5-2670 CPUs. |
| Software Dependencies | No | We implement all algorithms in Python, using the IPython environment (P erez and Granger, 2007) for parallelism. We compute effective sample size using RCODA (Plummer et al., 2006). The paper mentions Python and IPython, and RCODA, but does not provide specific version numbers for these software components. It also mentions Stan (version 1.0) but in the context of related work and not as part of their own implementation for the experiments presented. |
| Experiment Setup | Yes | In each experiment, we run each algorithm with 100 parallel chains. Unless otherwise noted, we burn in for 104 iterations and sample for 105 iterations. We run five trials for each experiment to estimate variability. A tuning period is used to adjust the MH step size so that the acceptance ratio is as close as possible to the value 0.234. We initialize each Markov chain from a spherical multivariate Gaussian centered on the origin. For each triple (D, C, K), we sample a D-dimensional multivariate Gaussian distribution centered on the origin... We initialize GESS from a broad spherical Gaussian distribution centered on the origin, and we run GESS for 500 seconds. The first half of the resulting samples are discarded, and the second half of the resulting samples are used to estimate the vector σ. |