Provably Fast Finite Particle Variants of SVGD via Virtual Particle Stochastic Approximation

Authors: Aniket Das, Dheeraj Nagaraj

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments (Section 8) show that GB-SVGD obtains similar performance as SVGD but requires fewer computations. 8 Experiments We compare the performance of GB-SVGD and SVGD. We take n = 100 and use the Laplace kernel with h = 1 for both. We pick the stepsize γ by a grid search for each algorithm. Additional details are presented in Appendix G. We observe that SVGD takes fewer iterations to converge, but the compute time for GB-SVGD is lower.
Researcher Affiliation Industry Aniket Das Google Research Bangalore, India ketd@google.com Dheeraj Nagaraj Google Research Bangalore, India dheerajnagaraj@google.com
Pseudocode Yes Algorithm 1 Virtual Particle SVGD (VP-SVGD) Algorithm 2 Global Batch SVGD (GB-SVGD)
Open Source Code No The paper does not provide an explicit statement or a link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes Sampling from Isotropic Gaussian (Figure 1): As a sanity check, we set π = N(0, I) with d = 5. Bayesian Logistic Regression (Figure 2) We consider the Covertype dataset which contains 580, 000 data points with d = 54. We consider the same priors suggested in Gershman et al. [16] and implemented in Liu and Wang [31].
Dataset Splits No The paper mentions using datasets but does not explicitly provide specific train, test, or validation split percentages, sample counts, or a detailed splitting methodology.
Hardware Specification No The paper does not explicitly describe specific hardware components such as GPU/CPU models, memory, or detailed computing environments used for running the experiments.
Software Dependencies No The paper does not specify particular software dependencies with version numbers (e.g., Python, specific libraries, or frameworks with their versions) that were used for the experiments.
Experiment Setup Yes We take n = 100 and use the Laplace kernel with h = 1 for both. We pick the stepsize γ by a grid search for each algorithm. We pick K = 10 for GB-SVGD. We take K = 40 for GB-SVGD. For both VP-SVGD and GB-SVGD, we use Ada Grad with momentum to set the step-sizes as per Liu and Wang [31]