A Heavy-Tailed Algebra for Probabilistic Programming

Authors: Feynman T. Liang, Liam Hodgkinson, Michael W. Mahoney

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical results confirm that inference algorithms that leverage our heavy-tailed algebra attain superior performance across a number of density modeling and variational inference (VI) tasks. 4 Experiments We now demonstrate that GGA-based density estimation yields improvements in tail estimation across several metrics.
Researcher Affiliation Academia Feynman Liang Department of Statistics University of California, Berkeley feynman@berkeley.edu Liam Hodgkinson School of Mathematics and Statistics University of Melbourne, Australia lhodgkinson@unimelb.edu.au Michael W. Mahoney ICSI, LBNL, and Department of Statistics University of California, Berkeley mmahoney@stat.berkeley.edu
Pseudocode Yes Algorithm 1: GGA tails static analysis pass Data: Abstract syntax tree for a PPL program Result: GGA parameter estimates for all random variables frontier [rv : Parents(rv) = ]; tails {}; while frontier = do next frontier.pop Left(); tails[next] compute GGA(next.op, next.parent); frontier frontier + next.children(); end return tails
Open Source Code Yes To illustrate an implementation of GGA for static analysis, we sketch the operation of the PPL compiler at a high-level and defer to the code in Supplementary Materials for details.
Open Datasets Yes super (superconductor critical temperature prediction dataset [23] with n = 256 and d = 154); who (life expectancy data from the World Health Organisation in the year 2013 [41] with n = 130, d = 18); air (air quality data [14] with n = 6941, d = 11); and blog (blog feedback prediction dataset [7] with n = 1024, d = 280).
Dataset Splits No All experiments are repeated for 100 trials, trained to convergence using the Adam optimizer with manually tuned learning rate. Additional details are available in Appendix D.
Hardware Specification No Our implementation target beanmachine [50] is a declarative PPL selected due to availability of a PPL compiler and support for static analysis plugins. Similar to [4, 46], it uses Py Torch [40] for GPU tensors and automatic differentiation.
Software Dependencies No To illustrate an implementation of GGA for static analysis, we sketch the operation of the PPL compiler at a high-level and defer to the code in Supplementary Materials for details. A probabilistic program is first inspected using Python s built-in ast module... it uses Py Torch [40] for GPU tensors and automatic differentiation. trained to convergence using the Adam optimizer with manually tuned learning rate.
Experiment Setup No All experiments are repeated for 100 trials, trained to convergence using the Adam optimizer with manually tuned learning rate. Additional details are available in Appendix D.