Variational Inference with Gaussian Score Matching

Authors: Chirag Modi, Robert Gower, Charles Margossian, Yuling Yao, David Blei, Lawrence Saul

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically compared GSM-VI to reparameterization BBVI on several classes of models, and with both synthetic and real-world data. We evaluate how well GSM-VI can approximate a variety of target posterior distributions.
Researcher Affiliation Collaboration Chirag Modi Center for Computational Astrophysics, Center for Computational Mathematics, Flatiron Institute, New York cmodi@flatironinstitute.org Charles C. Margossian Center for Computational Mathematics, Flatiron Institute, New York cmargossian@flatironinstitute.org Yuling Yao Center for Computational Mathematics, Flatiron Institute, New York yyao@flatironinstitute.org Robert M. Gower Center for Computational Mathematics, Flatiron Institute, New York rgower@flatironinstitute.org David M. Blei Department of Computer Science, Statistics, Columbia University, New York david.blei@columbia.edu Lawrence K. Saul Center for Computational Mathematics, Flatiron Institute, New York lsaul@flatironinstitute.org
Pseudocode Yes Algorithm 1: Gaussian Score Matching VI; Algorithm 2: Black-box variational inference
Open Source Code Yes We provide a Python implementation of GSM-VI algorithm at https://github.com/modichirag/GSM-VI.
Open Datasets Yes a collection of real-world Bayesian inference problems from the posterior DB database of datasets and models. M. Magnusson, P. Bürkner, and A. Vehtari. posteriordb: a set of posteriors for Bayesian inference and probabilistic programming, November 2022. URL https://github.com/ stan-dev/posteriordb.
Dataset Splits No The paper uses synthetic models where the true distribution is known, and for real-world models, it refers to the posteriordb. However, it does not provide explicit details on training, validation, and test dataset splits with percentages or counts for reproducibility.
Hardware Specification No The paper mentions running times but does not specify any hardware details like GPU/CPU models or specific machine configurations used for experiments.
Software Dependencies No The paper mentions 'Python implementation' and 'Jax', 'ADAM optimizer [19]', and 'bridgestan [6, 30]' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes The only free parameter in GSM-VI is the batch size B. ... In all studies of this section we report results for B = 2 and show that it is a good default baseline. We use the same batch size for BBVI. ... Unless specified otherwise, we initialize the variational approximation as a Gaussian distribution with zero mean and identity covariance matrix. ... We use the ADAM optimizer [19] with default settings but vary the learning rate between 10⁻¹ and 10⁻³.