/SlicStan

A Stan-like probabilistic programming language.

Primary LanguageF#MIT LicenseMIT

SlicStan

SlicStan1 is a Stan-like probabilistic programming language that translates to Stan. It provides automatic program transformations that allow for a more lightweight syntax and inference optimizations. There are three main ways in which SlicStan and Stan differ:

  1. SlicStan contains no program blocks, nor any annotations as to what block a variable belongs to (other than what the input data to the model is).
  2. In SlicStan, there is no need to distinguish between random variables defined using ~ (e.g. x ~ normal(0, 1)) and those defined using pseudo-random number generators (e.g. x = normal_rng(0, 1)).
  3. SlicStan supports discrete parameters, as long as the number of discrete parameters is fixed, known in advance, and their support is finite.

For example, the following program is a valid program in SlicStan:

real phi0 ~ beta(1, 1);
real theta0 ~ beta(1, 1); 

int<2> z1 ~ bernoulli(theta0);
real phi1 = phi0 * z1 + (1 - phi0) * (1 - z1);
data real y1 ~ normal(phi1, 1);

real theta1 = theta0 * z1 + (1 - theta0) * (1 - z1);
int<2> z2 ~ bernoulli(theta1);

This will translate to a Stan program with parameters phi0 and theta0, and generated quantities z1 and z2 (with z1 being automatically marginalized out from Stan's target density).

If you are interested in reading more about SlicStan and seeing examples, you can refer to either the 2019 POPL paper [1] (which also gives the operational density-based semantics of SlicStan), or the 2017 MSc thesis [2].

NOTE: SlicStan is a research repo, and as such the code is largely experimental, incomplete, and poorly documented. If you are looking for a reliable Bayesian workflow, please consider Stan. If you are interested in contributing SlicStan's or similar ideas to Stan, please have a look at the Stan3 repo.

1 SlicStan stands for "Slightly Less Intensely Constrained Stan".

Papers

[1] Gorinova, M. I., Gordon, A. D., Sutton, C., & Vákár, M. (2020). Conditional independence by typing. arXiv preprint arXiv:2010.11887.

[2] Gorinova, M. I., Gordon, A. D., & Sutton, C. (2019). Probabilistic programming with densities in SlicStan: efficient, flexible, and deterministic. Proceedings of the ACM on Programming Languages, 3(POPL), 1-30.

[3] Gorinova, M. I. (2017). Probabilistic Programming with SlicStan. MSc dissertation, University of Edinburgh.