Maddy Bowers

Maddy Bowers

MIT PhD candidate studying AI and program synthesis (they/she)

Last Updated: Sept 2022

Sept 2022: See the paper and code for our latest work Stitch: Top-Down Synthesis for Library Learning accepted to POPL 2023! See the GitHub linked above for a tutorial on using Stitch

I’m a PhD student at MIT co-advised by Armando Solar-Lezama in EECS and Josh Tenenbaum in BCS. My research is in combining methods from programming languages (PL) research with machine learning to tackle problems in artificial intelligence.

Research Interests

My research interests center on program synthesis and artificial intelligence. I’m particularly interested in neurosymbolic methods that bridge the machine learning and programming languages communities. I believe symbolic methods can augment neural methods to facilitate low-data learning, generalization, transfer learning, interpretability, and other desiderata.

I’m particularly interested in abstraction learning, as in Ellis et al.’s DreamCoder. I recently led a follow up work called Stitch (paper & code) which has been accepted to POPL 2023 and achieves a 1,000-10,000x speedup in abstraction learning over DreamCoder. I’m interested in exploring new applications of abstraction learning, and I’m particularly interested in its application to game-playing environments in world-model building and policy learning.

I previously published as Matthew Bowers.

Links

Publications