Today’s software systems increasingly involve learning. Learning is great: it lets you use data to generate programs when writing them by hand would be infeasible. However, learning has a price: learned systems can be opaque to understand and hard to control, leading to unexpected and undesirable system-level behavior. As our learned systems grow in scope and responsibility, it is becoming increasingly urgent that we tackle these challenges. In this talk we seek to answer the question: what is the right programming foundation for building robust, reliable, and safe systems that involve learned components?

We will break this problem up into two parts that are familiar to a programming languages audience: specification and implementation. I will argue that probabilistic specifications give us the right language for precisely describing nuanced notions like ”robustness” for learned systems. Then we ask: how can we implement a program that involves learning and verify that it meets a probabilistic specification? For this I will introduce probabilistic programming languages, languages that admit probability as a first-class construct. Unfortunately, we are still a long way away from this utopia of fully-verified end-to-end learned systems: I will discuss some of the open challenges and problems that keep us from getting there. Finally, I’ll conclude with a high-level discussion on other ways in which ML and PL interact.

I am an assistant professor at Northeastern University. My research focuses on programming languages, artificial intelligence, and machine learning. My goal is to design systems that make probabilistic modeling fast, accessible, and useful for solving every day reasoning tasks. Broadly my research focuses on (1) The design, implementation, and applications of probabilistic programming languages; (2) Foundations of probabilistic inference and tractable probabilistic modeling; (3) Automated reasoning and probabilistic verification.