What is Stan?

• An imperative probabilistic programming language
  • in the vein of BUGS, but can do arbitrary computation
• A Stan program defines a probability model
  • declares data and (constrained) parameter variables
  • all in order to define a log posterior density function
• Stan comes with a few inference algorithms
  • MCMC (HMC) for full Bayesian inference
  • VB for (one form of) approximate Bayesian inference
  • MLE for penalized maximum likelihood estimation

Why choose Stan?

• Expressive
  • Stan is a full imperative programming language
  • Easier to communicate models
• Robust
  • usually works; signals when it doesn’t
• Efficient
  • effective sample size per unit time winner
  • 3 forms of parallelism coming out this year (CPU, GPU, cluster)
• Great documentation, case studies, and community!

Why use Stan over lme4 et al?

• Bespoke models
• Propagate uncertainty
• Bayesian inference is often more intuitive
• Focus on better science

What’s next?

• Distributed likelihoods with multi-CPU parallelism
• Big matrix operations with GPU support
• Sparse matrix operations
• Distributed data: asynchronous expectation propagation
• Approximations: parallel max marginal mode