Soss: Declarative Probabilistic Programming via Runtime Code Generation

We present Soss, a declarative probabilistic programming language embedded in the Julia language. Soss represents statistical models in terms of abstract syntax trees, and uses staged compilation for on-demand generation of ``inference primitives'' (random sampling, log-density, etc) without requiring casual users to worry about such details.

The approach taken by Soss makes it easy to extend to take advantage of other packages in the rapidly-growing Julia ecosystem. At the time of this writing, Soss users can choose from several inference back-ends and connect easily with larger systems SymPy and Gen.