Skip to content

An embedded language for probabilistic programming and meta-programming.

License

Notifications You must be signed in to change notification settings

probcomp/metaprob

Repository files navigation

Metaprob

Build Status Stability: Experimental

A language for probabilistic programming and metaprogramming, embedded in Clojure.

Note: Metaprob is currently an unstable research prototype, with little documentation and low test coverage. Also, future versions may not be backwards compatible with this version. We do not recommend using it for any purpose other than basic research, and are not yet able to support users outside of the MIT Probabilistic Computing Project.

Key features

  • Models can be represented via generative code, i.e. ordinary code that makes stochastic choices
  • Models can also be represented via approximations, e.g. importance samplers with nontrivial weights
  • Custom inference algorithms can be written in user-space code, via reflective language constructs for:
    • tracing program executions
    • using partial traces to specify interventions and constraints
  • Generic inference algorithms are provided via user-space code in a standard library; adding new algorithms does not require modifying the language implementation
  • All Inference algorithms are ordinary generative code and can be traced and treated as models
  • New probability distributions and inference algorithms are first-class citizens that can be created dynamically during program execution

Motivations

  • Lightweight embeddings of probabilistic programming and inference metaprogramming
    • Interactive, browser-based data analysis tools (via ClojureScript)
    • Smart data pipelines suitable for enterprise deployment (via Clojure on the JVM)
  • “Small core” language potentially suitable for formal specification and verification
  • Teaching
    • Undergraduates and graduate students interested in implementing their own minimal PPL
    • Software engineers and data engineers interested in probabilistic modeling and inference
  • Research in artificial intelligence and cognitive science
    • Combining symbolic and probabilistic reasoning, e.g. via integration with Clojure’s core.logic
    • “Theory of mind” models, where an agent’s reasoning is modeled as an inference metaprogram acting on a generative model
    • Reinforcement learning and other “nested” applications of modeling and approximate inference
    • Causal reasoning, via a notion of interventions that extends Pearl's “do” operator
  • Research in probabilistic meta-programming, e.g. synthesis, reflection, runtime code generation

Modeling and tracing

Generative models are represented as ordinary functions that make stochastic choices.

;; Flip a fair coin n times
(def fair-coin-model
 (gen [n]
   (map (fn [i] (at i flip 0.5)) (range n))))
;; Flip a possibly weighted coin n times
(def biased-coin-model
 (gen [n]
   (let [p (at "p" uniform 0 1)]
     (map (fn [i] (at i flip p)) (range n)))))

Execution traces of models, which record the random choices they make, are first-class values that inference algorithms can manipulate.

We obtain scored traces using infer-and-score, which invokes a “tracing interpreter” that is itself a Metaprob program.

(infer-and-score :procedure fair-coin-model, :inputs [3])

Documentation