Variational Sparse Bayesian Quadrature (VSBQ) is a fast post-process Bayesian inference method for (potentially expensive) Bayesian models. It operates by recycling existing likelihood/density evaluations (e.g., from maximum-a-posteriori (MAP) optimization runs), fitting a regression surrogate (a sparse Gaussian process), and conducting variational inference to get a posterior approximation. Our current implementation is based on PyVBMC. benchflow
is a toolkit for running the benchmark experiments in the paper.
conda create -n vsbq python=3.9
conda activate vsbq
pip install -e ./benchflow
pip install -e ./pyvbmc
# Install the kernel for Jupyter
python -m ipykernel install --user --name vsbq
See the example notebook for a simple example of using VSBQ.
Please cite our paper if you find this work useful:
@misc{liFastPostprocessBayesian2024,
title = {Fast Post-Process {{Bayesian}} Inference with {{Variational Sparse Bayesian Quadrature}}},
author = {Li, Chengkun and Clart{\'e}, Gr{\'e}goire and Jørgensen, Martin and Acerbi, Luigi},
year = {2024},
number = {arXiv:2303.05263},
eprint = {2303.05263},
primaryclass = {cs, stat},
publisher = {arXiv},
doi = {10.48550/arXiv.2303.05263},
archiveprefix = {arxiv}
}