Probabilistic Programming: Foundations and Applications (2 day)

An intensive, two-day workshop on PROBABILISTIC PROGRAMMING, with contributed and invited talks, poster sessions, demos, and discussions.

Probabilistic models and inference algorithms have become standard tools for interpreting ambiguous, noisy data and building systems that learn from their experience. However, even simple probabilistic models can require significant effort and specialized expertise to develop and use, frequently involving custom mathematics, algorithm design and software development. State-of-the-art models from Bayesian statistics, artificial intelligence and cognitive science --- especially those involving distributions over infinite data structures, relational structures, worlds with unknown numbers of objects, rich causal simulations of physics and psychology, and the reasoning processes of other agents --- can be difficult to even specify formally, let alone in a machine-executable fashion.

PROBABILISTIC PROGRAMMING aims to close this gap, making variations on commonly-used probabilistic models far easier to develop and use, and pointing the way towards entirely new types of models and inference. The central idea is to represent probabilistic models using ideas from programming, including functional, imperative, and logic-based languages. Most probabilistic programming systems represent distributions algorithmically, in terms of a programming language plus primitives for stochastic choice; some even support inference over Turing-universal languages. Compared with representations of models in terms of their graphical-model structure, these representation languages are often significantly more flexible, but still support the development of general-purpose inference algorithms.

The workshop will cover, and welcomes submissions about, all aspects of probabilistic programming. Some questions of particular interest include:

1. What real-world problems can be solved with probabilistic programming systems today? How much problem-specific customization/optimization is needed? Where is general-purpose inference effective?

2. What does the probabilistic programming perspective, and in particular the representation of probabilistic models and inference procedures as algorithmic processes, reveal about the computability and complexity of Bayesian inference? When can theory guide the design and use of probabilistic programming systems?

3. How can we teach people to write probabilistic programs that work well, without having to teach them how to build an inference engine first? What programming styles support tractability of inference?

4. How can central ideas from software engineering --- including debuggers, validation tools, style checkers, program analyses, reusable libraries, and profilers --- help probabilistic programmers and modelers? Which of these tools can be built for probabilistic programs, or help us build probabilistic programming systems?

5. What new directions in AI, statistics, and cognitive science would be enabled if we could handle models that took hundreds or thousands of lines of probabilistic code to write?

Confirmed Keynote Speakers:

- Josh Tenenbaum (MIT)
- Stuart Russell (UC Berkeley)
- Christopher Bishop (Microsoft Research; University of Edinburgh)