Learning from Programs: Probabilistic Models, Program Analysis and Synthesis

摘要

The increased availability of massive codebases (e.g., GitHub) creates an exciting opportunity for new kinds of programming tools based on probabilistic models. Enabled by these models, tomorrow's tools will provide statistically likely solutions to programming tasks difficult or impossible to solve with traditional techniques. An example is our JSNice statistical program de-minification system (http://jsnice.org), now used by more than 150,000 users in every country worldwide. In this talk, I will discuss some of the latest developments in this new inter-disciplinary research direction: the theoretical foundations used to build probabilistic programming systems, the practical challenges such systems must address, and the conceptual connections between the areas of statistical learning, static analysis and program synthesis.