Involutive MCMC: a Unifying Framework

Authors: Kirill Neklyudov, Max Welling, Evgenii Egorov, Dmitry Vetrov

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 4, we demonstrate the potential of i MCMC formalism deriving irreversible counterparts of existent MCMC methods, and demonstrate empirical gains on different target distributions. We empirically validate these examples on a set of tasks, which includes a mixture of two 2d-Gaussians (Mo G2) and the posterior distribution of Bayesian logistic regression on several datasets (Australian, German, Heart) (see Appendix C.1 for details). For performance evaluation, we use the effective sample size (ESS), which measures how many independent samples the chain actually contains. In Table 2, we see that simply introducing irreversibility into the kernel may result in significant performance gains while having a negligible computational overhead.
Researcher Affiliation Collaboration Kirill Neklyudov 1 2 Max Welling 3 4 Evgenii Egorov 5 Dmitry Vetrov 1 2 1Samsung AI Center Moscow 2Samsung-HSE Laboratory, National Research University Higher School of Economics 3University of Amsterdam 4Canadian Institute for Advanced Research 5Skolkovo Institute of Science and Technology.
Pseudocode Yes We provide the pseudo-code in Algorithm 1 below. Algorithm 1 Involutive MCMC
Open Source Code Yes Code for reproducing the experiments is available at https://github.com/necludov/i MCMC.
Open Datasets Yes We empirically validate these examples on a set of tasks, which includes a mixture of two 2d-Gaussians (Mo G2) and the posterior distribution of Bayesian logistic regression on several datasets (Australian, German, Heart) (see Appendix C.1 for details). We consider four benchmark target distributions: synthetic Mo G2 and posterior distributions for Bayesian logistic regression on three real-world datasets from UCI Machine Learning Repository [Lichman, 2013]: Australian, German, and Heart.
Dataset Splits No The paper mentions using datasets for validation and performance evaluation, but it does not specify concrete details of training, validation, or test dataset splits (e.g., percentages, sample counts, or explicit splitting methodology) needed for reproduction.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes For all targets we choose α = 0.8. All the experiments were run for 20000 samples and repeated 100 times to provide robust results. To provide a robust comparison, we do not change the training process of the original algorithm. Moreover, we compare our modification against the original method, using the same already learned model as the proposal distribution.