Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Microcanonical Hamiltonian Monte Carlo

Authors: Jakob Robnik, G. Bruno De Luca, Eva Silverstein, Uroลก Seljak

JMLR 2023 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test our method on various benchmark problems in Section 3. The code with a tutorial is publicly available. Section 3 is titled "Experiments" and includes quantitative comparisons of MCHMC and MCLMC against NUTS HMC on several benchmark problems (Ill-conditioned Gaussian, Bi-modal distribution, Rosenbrock function, Neal's funnel, German credit, Stochastic Volatility, Cauchy distribution) using metrics like ESS.
Researcher Affiliation Academia Jakob Robnik: Physics Department University of California at Berkeley, G. Bruno De Luca: Stanford Institute for Theoretical Physics Stanford University, Eva Silverstein: Stanford Institute for Theoretical Physics Stanford University, Uroษ™s Seljak: Physics Department University of California at Berkeley and Lawrence Berkeley National Laboratory. All listed affiliations are academic institutions or associated national laboratories.
Pseudocode Yes Algorithm 1: MCHMC q = 0 algorithm. Algorithm 2: MCLMC q = 0 algorithm.
Open Source Code Yes The code with a tutorial is publicly available1. 1. https://github.com/JakobRobnik/MicroCanonicalHMC
Open Datasets Yes 3.5 German credit: This is a popular Bayesian regression test case (Dua and Graff, 2017). We use the model implementation from the Inference Gym (Sountsov et al., 2020) and initialize the sampler by a draw from a standard Gaussian, centered at the MAP solution.
Dataset Splits No The paper primarily focuses on evaluating the sampling efficiency and accuracy of MCHMC and MCLMC on various target distributions (many of which are synthetic or toy problems), rather than training/testing machine learning models with explicit dataset splits. While some real-world datasets like "German credit" and "S&P500 index" are used, specific train/validation/test splits are not provided; the evaluation is focused on posterior quality or sample properties.
Hardware Specification No The paper does not explicitly describe the hardware used for running its experiments. It focuses on the algorithmic details and performance metrics without mentioning specific CPU, GPU, or memory specifications.
Software Dependencies No The paper mentions the use of "NUTS (...) as implemented in the Num Pyro library (Phan et al., 2019)" and references "Blackjax: Library of samplers for jax (Lao and Louf, 2022)". However, specific version numbers for Num Pyro, Blackjax, or JAX are not provided.
Experiment Setup Yes We develop a fast tuning algorithm for bounce frequency (bounce strength for MCLMC) and the integration step-size. We do a short run with a few hundred steps and ฬ“ฬ“ฬ“0 = 0.5 to determine Var[E] and update the stepsize to ฬ“ฬ“ฬ“ = ฬ“ฬ“ฬ“0(0.0005 d/Var[E])1/4. We repeat this step a few times for convergence. ... For the 50D ICG example, we run 500 parallel chains, each initialized from a standard Gaussian. The integration stepsize is ฬ“ฬ“ฬ“ = 0.5.