A Stein variational Newton method
Authors: Gianluca Detommaso, Tiangang Cui, Youssef Marzouk, Alessio Spantini, Robert Scheichl
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We observe significant computational gains over the original SVGD algorithm in multiple test cases. We evaluate our new SVN method with the scaled Hessian kernel on a set of test cases drawn from various Bayesian inference tasks. |
| Researcher Affiliation | Academia | Gianluca Detommaso University of Bath & The Alan Turing Institute gd391@bath.ac.uk Tiangang Cui Monash University Tiangang.Cui@monash.edu Alessio Spantini Massachusetts Institute of Technology spantini@mit.edu Youssef Marzouk Massachusetts Institute of Technology ymarz@mit.edu Robert Scheichl Heidelberg University r.scheichl@uni-heidelberg.de |
| Pseudocode | Yes | The SVGD algorithm is summarised in Algorithm 1. The resulting Stein variational Newton method is summarised in Algorithm 2. |
| Open Source Code | Yes | Code and all numerical examples are collected in our Git Hub repository [1]. |
| Open Datasets | No | The paper uses simulated data generated based on specified prior distributions and forward operators (e.g., standard multivariate Gaussian prior for the double banana, Brownian motion prior for the conditioned diffusion), but does not provide concrete access information or links to established public datasets. |
| Dataset Splits | No | The paper describes experiments using particles initialized from a prior distribution and evolved over iterations to approximate a target distribution. However, it does not specify a training, validation, and test dataset split in the conventional sense for model reproduction. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9) that would be needed to replicate the experiments. |
| Experiment Setup | Yes | Here, the choice ε = 1 performs reasonably well in our numerical experiments. We run 10, 50, and 100 iterations of SVN-H. For discretization, we use an Euler-Maruyama scheme with step size Δt = 10^-2. We use β = 10 and T = 1. We take a single observation y = F(xtrue) + ξ, with xtrue being a random variable drawn from the prior and ξ ~ N(0, σ^2 I) with σ = 0.3. |