Probabilistic models, which more naturally quantify uncertainty when compared to their deterministic counterparts, are often difficult and tedious to implement. Probabilistic programming languages (PPLs) have greatly increased productivity of probabilistic modelers, allowing practitioners to focus on modeling, as opposed to the implementing algorithms for probabilistic (e.g. Bayesian) inference. Turing is a PPL developed entirely in Julia and is both expressive and fast due partly to Julia’s just-in-time (JIT) compiler being implemented in LLVM. Consequently, Turing has a more manageable code base and has the potential to be more extensible when compared to more established PPLs like STAN. One thing that may lead to the adoption of Turing is more benchmarks and feature comparisons of Turing to other mainstream PPLs. The aim of this project is to provide a more systematic approach to comparing execution times and features among several PPLs, including STAN, Pyro, nimble, and Tensorflow probability for a variety of Bayesian nonparametric (BNP) models, which are a class of models that provide a much modeling flexibility and often allow model complexity to increase with data size.

Organization

Student

Arthur Lui

Mentors

  • Hong Ge
  • Martin Trapp
  • Cameron Pfiffer
close

2020