Towards scaling up Markov chain Monte Carlo: an adaptive subsampling approach

Authors: Rémi Bardenet, Arnaud Doucet, Chris Holmes

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental All experiments were conducted using the empirical Bernstein-Serfling bound of Bardenet & Maillard (2013), which revealed equivalent to the empirical Bernstein bound in (9), and much tighter in our experience with MHSUBLHD than Hoeffding s bound in (7). All MCMC runs are adaptive Metropolis (Haario et al., 2001; Andrieu & Thoms, 2008) with target acceptance 25% when the di-mension is larger than 2 and 50% else (Roberts & Rosenthal, 2001). Hyperparameters of MHSUBLHD were set to p = 2, γ = 2, and δ = 0.01. The first two were found to work well with all experiments.
Researcher Affiliation Academia R emi Bardenet REMI.BARDENET@GMAIL.COM Arnaud Doucet DOUCET@STATS.OX.AC.UK Chris Holmes CHOLMES@STATS.OX.AC.UK Department of Statistics, University of Oxford, Oxford OX1 3TG, UK
Pseudocode Yes Figure 1. The pseudocode of the MH algorithm targeting the posterior π(θ) p(x1, ..., xn|θ)p(θ).
Open Source Code No The paper does not contain any statement about making source code publicly available or providing links to a code repository.
Open Datasets Yes We consider the dataset covtype.binary1 described in (Collobert et al., 2002). 1available at http://www.csie.ntu.edu.tw/ cjlin/libsvmtools/datasets/binary.html
Dataset Splits No The paper states 'we pick n = 400, 000 as a training set' for the covtype dataset and mentions a 'synthetic dataset of size n = 10^7', but it does not provide explicit training, validation, and test split percentages or counts needed for reproduction.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory specifications, or cloud instances) used to conduct the experiments.
Software Dependencies No The paper mentions specific statistical bounds and MCMC algorithms used (e.g., 'empirical Bernstein-Serfling bound', 'adaptive Metropolis'), but it does not specify any software dependencies with version numbers (e.g., programming languages, libraries, or solvers).
Experiment Setup Yes Hyperparameters of MHSUBLHD were set to p = 2, γ = 2, and δ = 0.01. The first two were found to work well with all experiments. We found empirically that the algorithm is very robust to the choice of δ.