Bayesian inference via sparse Hamiltonian flows

Authors: Naitong Chen, Zuheng Xu, Trevor Campbell

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we compare our method against other Hamiltonian-based VI methods and Bayesian coreset construction methods. Specifically, we compare the quality of posterior approximation, as well as the training and sampling times of sparse Hamiltonian flows (SHF), Hamiltonian importance sampling (HIS) [21], and unadjusted Hamiltonian annealing (UHA) [23] using real and synthetic datasets.
Researcher Affiliation Academia Naitong Chen Zuheng Xu Trevor Campbell Department of Statistics University of British Columbia [naitong.chen | zuheng.xu | trevor]@stat.ubc.ca
Pseudocode Yes Algorithm 1 Sparse Ham Flow
Open Source Code Yes Code is available at https://github.com/Naitong Chen/Sparse-Hamiltonian-Flows.
Open Datasets Yes The dataset that we use consists of N = 100, 000 flights...This dataset consists of airport data from https://www.transtats.bts.gov/DL_Select Fields.asp? gnoyr_VQ=FGJ and weather data from https://wunderground.com.
Dataset Splits No The paper describes the datasets used (e.g., flight data) but does not provide specific percentages or counts for training, validation, and test splits.
Hardware Specification Yes All experiments are performed on a machine with an Intel Core i7-12700H processor and 32GB memory.
Software Dependencies No The paper does not provide specific version numbers for software dependencies used in its implementation, such as programming languages or libraries (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes We initialize the weights to N/M (i.e., a uniform coreset), and select an initial step size for all dimensions. We use a warm start to initialize the parameters λr = (µr, Λr) of the quasi-refreshments... We obtain an unbiased estimate of the augmented ELBO gradient by applying automatic differentiation [38, 39] to the ELBO estimation function Algorithm 2, and optimize all parameters jointly using a gradient-based stochastic optimization technique such as SGD [40, 41] and ADAM [42]. Details of the experiments are in Appendix B.