Minibatch Gibbs Sampling on Large Graphical Models
Authors: Chris De Sa, Vincent Chen, Wing Wong
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Validation of MIN-Gibbs. Having characterized the effects of minibatching on Gibbs sampling, we present a synthetic scenario where Algorithm 2 can be applied. The Ising model (Ising, 1925) is a probabilistic model over an N N lattice, with domain x(i) { 1, 1}. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Cornell University, Ithaca, NY, USA 2Department of Statistics, Stanford University, Stanford, CA, USA. Correspondence to: Christopher De Sa <cdesa@cs.cornell.edu>. |
| Pseudocode | Yes | Algorithm 1 Gibbs sampling given: initial state x Ω loop sample variable index i uniformly from {1, . . . , n}... |
| Open Source Code | No | The paper does not provide any explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | No | The paper validates algorithms using synthetic scenarios (Ising and Potts models) where parameters are defined, rather than relying on external, publicly available datasets for which access information would be provided. For example: 'We chose an Ising model in which each site is fully connected (i.e. = N 2 1), and the strength of interaction Aij between any two sites is determined based on their distance by a Gaussian kernel.' |
| Dataset Splits | No | The paper uses synthetic models (Ising and Potts) and evaluates convergence over iterations, but does not define or use training, validation, or test dataset splits in the conventional sense for reproducibility. For example, it states: 'We started the algorithm with a unmixed configuration where each site takes on the same state (x(i) = 1 for all i).' |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments (e.g., CPU/GPU models, memory, or cloud instance types). |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers that would be needed to replicate the experiments. |
| Experiment Setup | Yes | We chose an Ising model in which each site is fully connected (i.e. = N 2 1), and the strength of interaction Aij between any two sites is determined based on their distance by a Gaussian kernel. We simulated Algorithm 2 on a graph with n = N 2 = 400 and inverse temperature β = 1. This β parameter was hand-tuned such that the Gibbs sampler seemed to mix in about the number of iterations we wanted to run. For this model, L = 2.21 and Ψ = 416.1. We started the algorithm with a unmixed configuration where each site takes on the same state (x(i) = 1 for all i). |