Challenges and Opportunities in High Dimensional Variational Inference

Authors: Akash Kumar Dhaka, Alejandro Catalina, Manushi Welandawe, Michael R. Andersen, Jonathan Huggins, Aki Vehtari

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our framework through an extensive empirical study using simulated data and many commonly used real datasets with both Gaussian and non-Gaussian target distributions.
Researcher Affiliation Collaboration Akash Kumar Dhaka* Aalto University, Silo AI akash.dhaka@aalto.fi
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The code for the experiments will be made available after acceptance using MIT license.
Open Datasets Yes We compare variational approximations for models and datasets from posteriordb in terms of accuracy of the estimated moments and predictive likelihood. https://github.com/stan-dev/posteriordb
Dataset Splits No We used an 80/20 training/test split on all datasets to compute the predictive likelihoods.
Hardware Specification No The experiments were carried on a laptop and an internal cluster with only CPU capability.
Software Dependencies Yes We use Stan [26] for model construction. [26] Stan Development Team. Stan modeling language users guide and reference manual. 2.26, 2020. URL https://mc-stan.org.
Experiment Setup Yes For stochastic optimization we use RMSProp with initial step size of 10 3 run for either Tmax iterations or until convergence was detected using a modified version of the algorithm by Dhaka et al. [6]. For the exclusive KL we use 10 draws for gradient estimation per iteration, while for the other divergences we use 200 draws, and a warm start at the solution of the exclusive KL. We use Tmax = 10,000.