Towards Understanding the Dynamics of Gaussian-Stein Variational Gradient Descent
Authors: Tianle Liu, Promit Ghosal, Krishnakumar Balasubramanian, Natesh Pillai
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conduct simulations to compare Gaussian SVGD dynamics with different kernels and the performance of the algorithms mentioned in the previous section. We consider three settings, Bayesian logistic regression, and Gaussian and Gaussian mixture targets. Here we present the results for Bayesian logistic regression as it involves a non-Gaussian but unimodal target and is one of the typical setups such that GVI is preferred in practice. |
| Researcher Affiliation | Academia | Tianle Liu Department of Statistics Harvard University Cambridge, MA 02138 tianleliu@fas.harvard.edu Promit Ghosal Department of Mathematics Massachusetts Institute of Technology Waltham, MA 02453 promit@mit.edu Krishnakumar Balasubramanian Department of Statistics University of California, Davis Davis, CA 95616 kbala@ucdavis.edu Natesh S. Pillai Department of Statistics Harvard University Cambridge, MA 02138 pillai@fas.harvard.edu |
| Pseudocode | Yes | Algorithm 1 Density-based Gaussian SVGD. ... Algorithm 2 Particle-based Gaussian SVGD. |
| Open Source Code | No | The paper does not provide an explicit statement about the release of source code or a link to a code repository. |
| Open Datasets | No | The paper describes generating data for Bayesian logistic regression and using Gaussian/Gaussian mixture targets, e.g., "Xi i.i.d. N(0, Id)", and "target is Gaussian N(µ, Σ) where µ Unif([0, 1]10)". However, it does not provide concrete access information (link, citation to a public dataset) for a publicly available dataset used for training. |
| Dataset Splits | No | The paper does not provide specific details about training, validation, or test dataset splits. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments. |
| Software Dependencies | No | The paper does not list specific software components with their version numbers. |
| Experiment Setup | Yes | The largest safe step sizes are 0.02, 0.1, 2, 0.8, 0.02, 0.2, 4, 4. ... In Figure 1a the same step size 0.01 is specified for all algorithms while for Figure 1b we choose the largest safe step size for each algorithm. |