Research

My research aims to automate and scale up principled probabilistic reasoning, in much the same way that tools like TensorFlow and PyTorch have automated and scaled up deep learning. To that end, I develop:

• Theoretical foundations for probabilistic programming: e.g., new semantic models for reasoning clearly about expressive probabilistic and differentiable programs, a type system for enforcing measure-theoretic correctness properties of inference algorithms, and a unifying framework for designing Monte Carlo and variational inference algorithms with powerful proposal distributions.

• Probabilistic program compilers that automate and speed up the math behind learning and inference: e.g., a new AD algorithm that automatically constructs gradient estimators for optimizing probabilistic objectives, and a new compiler for Gen programs that synthesizes fast, unbiased estimators of probability densities and their reciprocals, for use in Monte Carlo and variational inference.

• High-level tools that exploit this automation for scalable probabilistic reasoning in specific domains: e.g., a data-cleaning system that accurately and efficiently detects and corrects errors in real-world datasets with millions of records, and a new domain-specific language for steering large language models to behave more reliably.

Selected Publications

Selected publications are listed below; please see Google Scholar for a complete list. Asterisks (*) indicate co-first authorship.

• Probabilistic Programming with Programmable Variational Inference
  
  McCoy Becker*, Alexander Lew*, Xiaoyan Wang, Matin Ghavami, Mathieu Huot, Martin Rinard, Vikash Mansinghka
  PLDI 2024
• ωPAP Semantics: Reasoning Denotationally About Higher-Order, Recursive Probabilistic and Differentiable Programs
  
  Mathieu Huot*, Alexander Lew*, Vikash Mansinghka, Sam Staton
  LICS 2023, distinguished paper
• Probabilistic Programming with Stochastic Probabilities
  
  Alexander Lew, Matin Ghavami, Martin Rinard, Vikash Mansinghka
  PLDI 2023
• ADEV: Sound Automatic Differentiation of Expected Values of Probabilistic Programs
  
  Alexander Lew*, Mathieu Huot*, Sam Staton, Vikash Mansinghka
  POPL 2023, distinguished paper
• SMCP³: Sequential Monte Carlo with Probabilistic Program Proposals
  
  Alexander Lew*, George Matheos*, Tan Zhi-Xuan, Matin Ghavami, Nishad Gothoskar, Stuart Russell, Vikash Mansinghka
  AISTATS 2023
• Recursive Monte Carlo and Variational Inference with Auxiliary Variables
  
  Alexander Lew, Marco Cusumano-Towner, Vikash Mansinghka
  UAI 2022
• PClean: Bayesian Data Cleaning at Scale with Domain-Specific Probabilistic Programming
  
  Alexander Lew, Monica Agrawal, David Sontag, Vikash Mansinghka
  AISTATS 2021, oral presentation
• Leveraging Unstructured Statistical Knowledge in a Probabilistic Language of Thought
  
  Alexander Lew, Michael Henry Tessler, Vikash Mansinghka, Joshua Tenenbaum
  COGSCI 2020
• Trace Types and Denotational Semantics for Sound Programmable Inference in Probabilistic Languages
  
  Alexander Lew, Marco Cusumano-Towner, Benjamin Sherman, Michael Carbin, Vikash Mansinghka
  POPL 2020
• Few-Shot Bayesian Imitation Learning with Logical Program Policies
  
  Tom Silver, Kelsey Allen, Alexander Lew, Leslie Kaelbling, Josh Tenenbaum
  AAAI 2020
• Gen: A General-Purpose Probabilistic Programming System with Programmable Inference
  
  Marco Cusumano-Towner, Feras Saad, Alexander Lew, Vikash Mansinghka
  PLDI 2019

Recorded Talks

Some of my research talks are available online with video recordings:

• Integrating Language into Intelligent Architectures (with Lio Wong)
  
  2023-08-17 / Simons Institute Workshop on Large Language Models and Transformers
• Probabilistic Programming with Stochastic Probabilities
  
  2023-06-20 / PLDI 2023
• ADEV: Sound Automatic Differentiation of Expected Values of Probabilistic Programs
  
  2023-01-19 / POPL 2023
• What do the posteriors of probabilistic programs look like?
  
  2023-01-15 / Languages for Inference Workshop 2023
• Towards Denotational Semantics of AD for Higher-Order, Recursive, Probabilistic Languages
  
  2022-01-16 / Languages for Inference Workshop 2022
• Reasoning about AD in Higher-Order, Recursive, Probabilistic Languages (best talk award)
  
  2021-12-13 / NeurIPS Differentiable Programming Workshop 2021
• PClean: Bayesian Data Cleaning at Scale with Domain-Specific Probabilistic Programming
  
  2021-04-14 / AISTATS 2021
• Probabilistic Programming with Correct-by-Construction Densities
  
  2021-02-18 / Joint PPS-PIHOC-DIAPASoN Workshop 2021
• On Automating the Estimation of Radon-Nikodym Derivatives for Probabilistic Programs
  
  2021-01-17 / Languages for Inference Workshop 2021
• Leveraging Unstructured Statistical Knowledge in a Probabilistic Language of Thought
  
  2020-08-03 / COGSCI 2020
• Trace Types and Denotational Semantics for Sound Programmable Inference in Probabilistic Languages
  
  2020-01-23 / POPL 2020
• Probabilistic Scripts for Automating Common-Sense Tasks
  
  2019-09-14 / StrangeLoop 2019
• Gen: A General-Purpose Probabilistic Programming System (delivered in Marco Cusumano-Towner’s stead)
  
  2019-07-25 / JuliaCon 2019

Teaching

At MIT, I co-taught a January-term course on applied probabilistic programming, and served twice as a TA for 6.885: Probabilistic Programming & AI.

From 2015 to 2018, I taught computer science full-time at Commonwealth School.

• CS 1: Intro to Program Design
• CS 2: AP Computer Science Principles
• CS 3: Data Structures and Algorithms
• CS 4: Advanced Topics in Computer Science

From 2015 to 2019, I served as a TA at the Duke Machine Learning Summer School.

• TensorFlow fundamentals

Contact

Email me at alexlew AT mit DOT edu.