Feras Saad
I am an Assistant Professor in the Computer Science Department at Carnegie Mellon, affiliated with the Principles of Programming and Artificial Intelligence groups.
The goal of my research is to develop scalable computing systems that enable flexible probabilistic modeling and sound probabilistic inference. This work integrates ideas from programming languages and AI, which together give us the tools to engineer principled probabilistic reasoning systems with a high degree of automation, accuracy, and scale.
My research interests span the computing stack---from foundational questions (e.g., how to represent probability on finite computers) to programming systems (languages, semantics, compilers) up to statistical models and algorithms for challenging data science applications. I also maintain open-source software libraries to help practitioners use these techniques in their own problems. See research and software for an overview.
Spring 2024: I am organizing the PROBPROG Seminar Series. Submissions are welcome.
Background
I received the PhD, MEng, and SB degrees in Computer Science from MIT. My theses on probabilistic programming were recognized with the Sprowls PhD Thesis Award in AI+Decision Making and Johnson MEng Thesis Award in Computer Science. Prior to CMU, I was a Visiting Research Scientist at Google.
Contact
|
|
Research Opportunities
I am actively recruiting students and postdocs with strong quantitative backgrounds.
- Prospective Students: If you are interested in working with me as a graduate student, please apply to the CMU CS PhD program and mention my name in your application.
- Current CMU Students: Please send me an email with your specific research interests, transcript, and CV. For SCS undergrads, important background includes 15-259, 10-315 or 15-281, 15-210, which are typically completed by the end of sophomore year.
- Postdocs: Please send me an email with your CV, two representative papers, and a description of your research interests.
Probabilistic Programming.
Researchers in diverse areas (e.g., econometrics, cognitive science,
bioinformatics) routinely develop computational models of uncertain data
and perform inference from observed data to learn about the underlying
data-generating process. Probabilistic programs, which produce outputs by
making random choices, are a unifying and uniquely expressive approach to
modeling and inference over these processes. These languages
aim to rigorously formalize the
semantics of probabilistic models, automate the process of discovering
models for complex data, and scale-up the hard aspects of
probabilistic inference in high-dimensional settings.
[
Automated Probabilistic Model Discovery.
A grand challenge in AI is automating the process of
discovering accurate and interpretable probabilistic models for reasoning about data.
I am interested in studying this problem through the lens of probabilistic program synthesis,
formalized as Bayesian structure learning over rich generative model families.
[
Statistical Estimators & Tests.
As probabilistic programs become a more widespread representation,
new statistical techniques are needed for inferring their properties
through the black-box computational inferences they expose. While static or
analytical analysis is sometimes possible, a more flexible
approach is to infer program properties using
simulation-based statistical analyses of sampled execution traces.
[
Random Variate Generation.
Generating random variables is a fundamental operation in scientific computation.
This work explores (1) computational and information-theoretic limits
of random sampling algorithms;
and (2) developing new samplers that are
theoretically optimal or near-optimal (in entropy consumption and/or statistical
sampling error) while also extremely efficient in practice.
[
Software + Applications.
A central motivation of my work is to develop performant, freely available
