Approximation Based Variance Reduction for Reparameterization Gradients

Authors: Tomas Geffner, Justin Domke

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically show that this control variate leads to large improvements in gradient variance and optimization convergence for inference with non-factorized variational distributions.
Researcher Affiliation Academia Tomas Geffner College of Information and Computer Science University of Massachusetts, Amherst tgeffner@cs.umass.edu Justin Domke College of Information and Computer Science University of Massachusetts, Amherst domke@cs.umass.edu
Pseudocode Yes Algorithm 1 SGVI with the proposed control variate.
Open Source Code No The paper does not provide an explicit statement or link for open-source code for the methodology.
Open Datasets Yes We use three different models: Logistic regression with the a1a dataset, hierarchical regression with the frisk dataset [7], and a Bayesian neural network with the red wine dataset. The latter two are the ones used by Miller et al. [17].
Dataset Splits No The paper does not explicitly provide training/validation/test dataset splits needed to reproduce the experiment.
Hardware Specification Yes we use Py Torch 1.1.0 on an Intel i5 2.3GHz
Software Dependencies Yes we use Py Torch 1.1.0 on an Intel i5 2.3GHz
Experiment Setup Yes We use Adam [13] to optimize the parameters w of the variational distribution qw (with step sizes between 10 5 and 10 2). We use Adam with a step size of 0.01 to optimize the parameters v of the control variate, by minimizing the proxy to the variance from Eq. 12. We parameterize Bv as a diagonal plus rank-rv. We set rv = 10 when diagonal or diagonal plus low rank variational distributions are used, and rv = 20 when a full-rank variational distribution is used. We use M = 10 and M = 50 samples from qw to estimate gradients.