Practical Hamiltonian Monte Carlo on Riemannian Manifolds via Relativity Theory

Authors: Kai Xu, Hong Ge

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We implement all samplers studied in this paper using Advanced HMC.jl (Xu et al., 2020).3 Derivative implementation in (5.3) is tested by finite differentiation. Geweke tests (Geweke, 2004; Grosse & Duvenaud, 2014) are used to validate the correctness of samplers (detailed in appendix C). Appendix D lists default hyper-parameters used across experiments.
Researcher Affiliation Collaboration 1MIT-IBM Watson AI Lab, Cambridge MA, United States 2University of Cambridge, Cambridge, United Kingdom.
Pseudocode Yes Algorithm 1 Momentum sampling via Box Muller transform
Open Source Code Yes Available at https://github.com/Turing Lang/ Advanced HMC.jl
Open Datasets Yes We use the 2-dimensional Neal s funnel (Neal, 2011), hierarchical Bayesian logistic regression (HBLR) model and a log-Gaussian Cox point process (log GCPP) model, that are previously used to benchmark HMC algorithms (Heng & Jacob, 2019; Xu et al., 2021). The HBLR problem has a dimensionality of 26, and the log GCPP problem has a dimensionality of 64.
Dataset Splits No The paper does not provide specific dataset split information (e.g., train/validation/test percentages or counts) for the experiments.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions software like 'Advanced HMC.jl', 'MCMCDebugging.jl', 'Turing', 'Zygote.jl', and 'Forward Diff.jl', but does not provide specific version numbers for these dependencies.
Experiment Setup Yes Appendix D lists default hyper-parameters used across experiments. Table 4: Common hyper-parameters used across experiments in section 6. parameter name value comment the number of leapfrog steps 8 the number of fixed-point iterations 6 scale of identity matrix added to Hessian (λ) 1 10 2 we use H + λI to regularize the Hessian initial position distribution U( 1, 1) this follows Betancourt (2013)