Fast Conditional Mixing of MCMC Algorithms for Non-log-concave Distributions

Authors: Xiang Cheng, Bohan Wang, Jingzhao Zhang, Yusong Zhu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct experiments to verify the theoretical results and compare global mixing versus conditional mixing for Gaussian mixture models. We take three Gaussian mixtures: ν1 = 0.9N1( 10, 1) + 0.1N1(10, 1), ν2 = 0.15N1( 5, 1) + 0.15N1( 2.5, 1) + 0.3N1(0, 1) + 0.2N1(2.5, 1) + 0.2N1(5, 1), and ν3 = 0.4N2(( 5, 5), I2) + 0.4N2((5, 5), I2) + 0.1N2(( 5, 5), I2) + 0.1N2((5, 5), I2) as our target distributions. We use Algorithm 1 as our sampling algorithm, and set step size h = 10 2. The initial distributions are both uniform in a large enough range. We plot the sampling distribution after T = 500, 5000, 500 rounds respectively in Figure 1a, 1b, and 1c, and plot the conditional and global KL divergence in Figure 1d, 1e, and 1f.
Researcher Affiliation Academia Xiang Cheng MIT x.cheng@berkeley.edu Bohan Wang USTC bhwangfy@gmail.com Jingzhao Zhang IIIS, Tsinghua; Shanghai Qizhi Institute jingzhaoz@mail.tsinghua.edu.cn Yusong Zhu Tsinghua University zhuys19@mails.tsinghua.edu.cn
Pseudocode Yes Algorithm 1 Langevin Monte Carlo Input: Initial parameter z, potential function V , step size h, number of iteration T 1: Initialization z0 z 2: For t = 0 T: 3: Generate Gaussian random vector ξt N(0, Id) 4: Update z(t+1)h zth h V (zth) + 2hξt 5: End For
Open Source Code No The paper does not provide an explicit statement or link to open-source code for the described methodology.
Open Datasets No The paper uses constructed target distributions (mixtures of Gaussians) for its experiments, rather than external publicly available datasets. For instance, in Section 6.1, it states: 'We take three Gaussian mixtures: ν1 = 0.9N1( 10, 1) + 0.1N1(10, 1), ν2 = 0.15N1( 5, 1) + 0.15N1( 2.5, 1) + 0.3N1(0, 1) + 0.2N1(2.5, 1) + 0.2N1(5, 1), and ν3 = 0.4N2(( 5, 5), I2) + 0.4N2((5, 5), I2) + 0.1N2(( 5, 5), I2) + 0.1N2((5, 5), I2) as our target distributions.'
Dataset Splits No The paper describes experiments involving sampling from target distributions rather than training models on datasets with explicit train/validation/test splits. Therefore, it does not specify dataset splits for validation.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as CPU or GPU models, or memory specifications.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies, libraries, or programming languages used in the experiments.
Experiment Setup Yes We use Algorithm 1 as our sampling algorithm, and set step size h = 10 2. The initial distributions are both uniform in a large enough range. We plot the sampling distribution after T = 500, 5000, 500 rounds respectively in Figure 1a, 1b, and 1c, and plot the conditional and global KL divergence in Figure 1d, 1e, and 1f.