Particle-Gibbs Sampling for Bayesian Feature Allocation Models

Authors: Alexandre Bouchard-Côté, Andrew Roth

JMLR 2021 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare the performance of our proposed methods to the standard Gibbs sampler using synthetic and real data from a range of feature allocation models. Our results suggest that row wise updates using the PG methodology can significantly improve the performance of samplers for feature allocation models.
Researcher Affiliation Academia Alexandre Bouchard-Côt e EMAIL Department of Statistics, University of British Columbia. Andrew Roth EMAIL Department of Computer Science, University of British Columbia
Pseudocode Yes Algorithm 1 Sample a row of the feature allocation using the element wise Gibbs update. Algorithm 2 Sample a row of the feature allocation using the row wise Gibbs update. Algorithm 3 Sample a row of the feature allocation using the particle Gibbs update. Algorithm 4 Conditional resampling for DPF. Algorithm 5 Sample a row of the feature allocation using the discrete particle filter update.
Open Source Code Yes Code implementing the samplers and models is available online at https://github. com/aroth85/pgfa. All experiments were done using version 0.3.4 of the software. Code for performing the experiments is available online at https://github.com/aroth85/pgfa_ experiments.
Open Datasets Yes We downloaded the Yale Face data from http://vision.ucsd.edu/datasets/yale_face_ dataset_original/yalefaces.zip.
Dataset Splits Yes We randomly assigned 10% of the data matrix to be missing and used these entries to compute root mean square reconstruction error (RMSE).
Hardware Specification No The paper does not explicitly describe the specific hardware used (e.g., CPU, GPU models, memory, or cloud instances) for running the experiments. It only states that 'All experiments were done using version 0.3.4 of the software.'
Software Dependencies Yes All experiments were done using version 0.3.4 of the software.
Experiment Setup Yes Based on these results we used the following parameter values for subsequent experiments. Annealing power: 1.0, Number of particles: 20, Resample threshold: 0.5, Particle Gibbs resampling scheme: Multinomial, Test path: Zeros.