Heavily inspired by Gen.jl and Poirot.jl: Jaynes is a minimal trace-based PPL but includes the usage of IR manipulations for non-standard interpretation and analysis, which may help by providing information which can be used during inference programming.
This might allow us to do cool things like:
- Grab the dependency graph of a probabilistic program as a static pass and analyze it!
- Possibly store analysis meta-data for inference programming (i.e. sub-graphs with exponential conjugacy can be identified).
- Belief propagation and trace-based inference in one PPL like whaaaaaaaaaa
Work in progress :)