Probabilistic programming with programmable inference

Abstract

We introduce inference metaprogramming for probabilistic programming languages, including new language constructs, a formalism, and the rst demonstration of e ectiveness in practice. Instead of relying on rigid black-box inference algorithms hard-coded into the language implementation as in previous probabilistic programming languages, infer- ence metaprogramming enables developers to 1) dynamically decompose inference problems into subproblems, 2) apply in- ference tactics to subproblems, 3) alternate between incorpo- rating new data and performing inference over existing data, and 4) explore multiple execution traces of the probabilis- tic program at once. Implemented tactics include gradient- based optimization, Markov chain Monte Carlo, variational inference, and sequental Monte Carlo techniques. Inference metaprogramming enables the concise expression of proba- bilistic models and inference algorithms across diverse elds, such as computer vision, data science, and robotics, within a single probabilistic programming language.

Building machines that learn and think like people

“…On the contrary, some of the most exciting recent progress has been in new forms of probabilistic machine learning (Ghahramani 2015). For example, researchers have developed automated statistical reasoning techniques (Lloyd et al 2014), automated techniques for model building and selection (Grosse et al 2012), and probabilistic programming languages (e.g., Gelman et al 2015; Goodman et al 2008; Mansinghka et al 2014). We believe that these approaches will play important roles in future AI systems, and they are at least as compatible with the ideas from cognitive science we discuss here.…”

Sound probabilistic inference via guide types

“…These inference algorithms usually require some guide programs, which can have a substantial influence on the performance of the inference. Although many PPLs provide mechanisms for automatically generating those guide programs, the ability to allow users to customize them, has been shown to be helpful, and sometimes crucial, for effective inference [8,17,19,41]. However, customizability introduces non-trivial challenges to ensuring soundness of Bayesian inference.…”

“…It has applications in many fields, including artificial intelligence [22], cognitive science [28], and applied statistics [21]. Because there is not a single known inference algorithm that works well for all models [41], several PPLs have recently added support for programmable inference [8,17,20,41,46,63]. This capability allows users to customize inference algorithms based on the characteristics of a particular model or dataset.…”

“…This capability allows users to customize inference algorithms based on the characteristics of a particular model or dataset. Researchers have shown that programmable inference enables improved inference performance on a variety of modeling problems [8,17,19,41].…”

“…In our development, we follow a common scheme of tracebased programmable inference that underlies Pyro [8], Venture [41], Gen [17], etc. These PPLs define the meaning of a probabilistic program by a probability distribution on sample traces that record all the random samples that the program draws during its execution.…”