Soss - Lightweight Probabilistic Programming with No Performance Ceiling
summary
Concrete claim that Soss encodes models as manipulable Julia code and removes the usual Python/C++ split, demonstrated by a working NUTS sampler, a compiler-produced log-density generator, model combinators and an LDA dependency graph. Back ends are pluggable so performance is limited only by chosen backend, not the framework. Progress looks real but early expect prototype maturity rather than a full ecosystem. back-end agnostic
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:08Why probabilistic programming matters
- 3:20Existing systems
- 7:18Two-language problem
- 9:03Soss approach
- 13:04Model combinators
- 16:21Compiling to log density
- 18:05Dependency graphs
- 19:59Future directions
- 21:44Combinators and variational inference
- 22:25JuliaCon 2019
related talks
36mProbabilistic Programs Which Make (Common) Sense
47mThe State of the Art for Probabilistic Programming
18mState Space Model Programming in Turing.jl
32mProbabilistic and Differentiable Programming in Scientific Simulators
10mJuliaBUGS: A Graph-Based Probabilistic Programming Language using BUGS syntax
29mMotivating Probabilistic Programming
01h 03mAn Overview of Probabilistic Programming
33mScala DSLs and Probabilistic Programming