Diffusive Gibbs Sampling
Authors: Wenlin Chen, Mingtian Zhang, Brooks Paige, José Miguel Hernández-Lobato, David Barber
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Di GS on three complex multi-modal sampling tasks across various domains: a mixture of 40 Gaussians, Bayesian neural network, and molecular dynamics. Table 2 shows the Maximum Mean Discrepancy (MMD) (Gretton et al., 2012) (computed with 5 kernels with bandwidths {2 2, 2 1, 20, 21, 22}) between the true samples and samples generated by each sampler and the Mean Absolute Error (MAE) between the true and estimated expectations of a quadratic function under the Mo G-40 target. This demonstrates that our method significantly outperforms all baselines on this problem. |
| Researcher Affiliation | Academia | 1University of Cambridge, Cambridge, UK 2Max Planck Institute for Intelligent Systems, T ubingen, Germany 3University College London, London, UK. |
| Pseudocode | Yes | Algorithm 1 summarizes the proposed Diffusion Gibbs Sampling (Di GS) procedure. |
| Open Source Code | Yes | The code of our experiments can be found in https:// github.com/Wenlin-Chen/Di GS. |
| Open Datasets | Yes | We first consider a synthetic problem from Midgley et al. (2023), which is a 2D Mo G with 40 mixture components. We sample the ground-truth parameters θ p(θ) from the prior and use θ to generate N = 500 training points and 500 test points for evaluation. The PT samples are taken from Midgley et al. (2023), which is generated using 21 chains starting at temperature 300K... We follow Midgley et al. (2023) and treat 107 PT samples as the ground-truth. |
| Dataset Splits | No | We sample the ground-truth parameters θ p(θ) from the prior and use θ to generate N = 500 training points and 500 test points for evaluation. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) are mentioned in the paper. |
| Software Dependencies | No | The paper mentions various algorithms (MALA, HMC, PT, ULA) but does not provide specific version numbers for any software dependencies or libraries used in the implementation. |
| Experiment Setup | Yes | MALA runs 1,000 Langevin steps per sample with a step size of 1 10 1. HMC runs 1,000 leapfrog steps per sample with a step size of 1 10 1. Di GS uses T = 1 noise level with α = 0.1 and σ2 = 1 α2, 200 Gibbs sweeps, and 5 MALA denoising sampling steps per Gibbs sweep with a step size of 1 10 1. |