[Pl-seminar] Talk Thursday at MSR New England: Self-applicable probabilistic inference without interpretive overhead

[Aaron Turon]

Self-applicable probabilistic inference without interpretive overhead
Chung-chieh Shan, Rutgers University (joint work with Oleg Kiselyov)

Thursday February 11, 4pm
Microsoft Research New England (1 Memorial Drive, Cambridge)
(Directions: Take the elevator from the ground floor up to the first
floor, then go past the reception desk and to the left.  There will be
a sign reading "Microsoft Research New England" to indicate where to
go, and the receptionist can help if there are any questions.)

Probabilistic programming is an expressive way to build stochastic
models and inference procedures as separate reusable modules.  We
express models and inference alongside each other as ordinary code in
the same general-purpose language.  This way, deterministic parts of
models run at full speed, so inference procedures can reason about
themselves without interpretive overhead.  This ability lets us model
bounded-rational theories of mind.

We use existing facilities of the language, such as rich libraries,
optimizing compilers, and types, to develop realistic models whose
inference performance is competitive with the state of the art.  In
particular, a wide range of models can be expressed using memoization,
and we introduce a new, general algorithm for importance sampling with
look-ahead.