Bregman Proximal Langevin Monte Carlo via Bregman-Moreau Envelopes

Authors: Tim Tsz-Kit Lau, Han Liu

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform numerical experiments of sampling anisotropic Laplace distributions which have nonsmooth potentials. Other additional numerical experiments are given in Appendix D. In this section, we use bold lower case letters θ = (θi) 1 i d Rd to denote vectors. All numerical implementations can be found at https://github.com/ timlautk/bregman_prox_langevin_mc. ... The marginal empirical densities are given in Figure 1 (Figure D.4 for BMMMLA in Appendix B).
Researcher Affiliation Academia 1Department of Statistics and Data Science, Northwestern University, Evanston, IL, USA 2Department of Computer Science, Northwestern University, Evanston, IL, USA.
Pseudocode Yes Algorithm 1 The Bregman Moreau Mirrorless Mirror-Langevin Algorithm (BMMMLA)
Open Source Code Yes All numerical implementations can be found at https://github.com/ timlautk/bregman_prox_langevin_mc.
Open Datasets Yes For such a nonsmooth sampling task, inspired by Vorstrup Goldman et al. (2021); Bouchard-Cˆot e et al. (2018), we consider the case where f = 0 and g(θ) = α θ 1 = Pd i=1 αi|θi| with α = (1, 2, . . . , d) . ... We take d = 100, N = 1000 and θ = (0 10, 0.1 1 10, 0.2 1 10, . . . , 0.9 1 10) R100 as the ground truth. Then, each xn,i is generated from a standard Gaussian distribution and each yn is sampled following (D.5) with θ = θ .
Dataset Splits No The paper focuses on sampling algorithms and does not describe experiments with traditional training, validation, and test splits of a fixed dataset. Instead, it generates samples and evaluates their properties against a known distribution or estimated posteriors.
Hardware Specification No The paper does not specify any hardware details like CPU or GPU models used for the experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies used.
Experiment Setup Yes We consider d = 100, draw K = 105 samples, with a tight Bregman Moreau envelope using a small smoothing parameter λ = 10 5 and a small step size γ = λ/2. The parameter of the hyperbolic entropy is β = (2 d i + 1) 1 i d. ... In the experiment, we consider the case where ai = i and bi = i for all i Jd K. We use γ = 0.01, λ = 1 and β = (2 d i + 1) 1 i d. ... In addition, we choose α1 = (10 1 10, 9 1 10, . . . , 1 1 10) and α2 = 0.1. Again, we use the hypentropy functions ϕβ (for the mirror map) and ψσ (for the Bregman Moreau envelope), with β = (2i 1/4 1 10) 1 i 10 and σ = (α2 1,i) 1 i d. We also use a step size γ = 5 10 4 and a smoothing parameter λ = 0.01.