On the properties of variational approximations of Gibbs posteriors

Authors: Pierre Alquier, James Ridgway, Nicolas Chopin

JMLR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We specialize our results to several learning tasks (classification, ranking, matrix completion), discuss how to implement a variational approximation in each case, and illustrate the good properties of said approximation on real datasets. ... We now compare the numerical performance of the mean field and full covariance VB approximations to the Gibbs posterior (as approximated by SMC, see Section 3.1) for the classification of standard datasets; see Table 1.
Researcher Affiliation Academia Pierre Alquier EMAIL James Ridgway EMAIL Nicolas Chopin EMAIL ENSAE 3 Avenue Pierre Larousse 92245 MALAKOFF, FRANCE
Pseudocode Yes The pseudo-code below is given for an adaptive sequence of temperatures. Algorithm 1 Tempering SMC ... Algorithm 2 Systematic resampling ... Algorithm 3 Deterministic annealing ... Algorithm 4 Stochastic Gradient Descent
Open Source Code Yes We also provide a R package1, written in C++ to compute a Gaussian variational approximation in the case of the hinge risk. 1. PACVB package: https://cran.r-project.org/web/packages/PACVB/index.html
Open Datasets Yes The datasets are all available in the UCI repository3 except for the DNA dataset which is part of the R package mlbench by Leisch and Dimitriadou (2010). 3. https://archive.ics.uci.edu/ml/datasets.html
Dataset Splits Yes When no split between the training sample is provided we split the data in half.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or memory specifications used for running the experiments. It only mentions general aspects of numerical performance and convergence speed.
Software Dependencies No We also provide a R package1, written in C++ to compute a Gaussian variational approximation in the case of the hinge risk. While it mentions R and C++, no specific version numbers for these languages or any libraries are provided.
Experiment Setup Yes The hyperparameters are chosen by cross-validation. ... Stochastic VB with fixed temperature λ = 100 for Pima and λ = 1000 for adult. ... In all our experiments we take c = 1 and η = 0.9.