Neurosymbolic programming for science

A 2022 position paper, presented at the NeurIPS 2022 AI for Science workshop, on combining neural learning with symbolic programs. It aims to give researchers interpretable machine-learning tools.

“Neurosymbolic programming for science” is a position paper by Jennifer J. Sun, Megan Tjandrasuwita, Atharva Sehgal, Armando Solar-Lezama, Swarat Chaudhuri, Yisong Yue, and Omar Costilla-Reyes. It was posted to arXiv in October 2022 and presented at the NeurIPS 2022 AI for Science workshop. Dr. Costilla Reyes described it as his first research paper as a senior researcher at MIT CSAIL, where he served as research scientist and research project manager for the multi-institution expedition behind it. A position paper argues for a research direction and sets out the open problems that direction addresses.

Neural networks learn complex patterns from data but are often difficult to interpret, while symbolic programs are explicit, readable procedures built from concepts and rules. Neurosymbolic programming combines the two as a way to blend learned representations with symbolic domain knowledge from scientists, including prior knowledge, experimental context, high-level concepts, and known constraints, so that outputs can remain interpretable to researchers. The paper identifies opportunities and open challenges for applying it across scientific workflows, using behavior analysis as a real-world example. Its figure maps the neurosymbolic programming lifecycle, from data curation and domain knowledge through model training to evaluation and deployment, onto the stages of the scientific process, from choosing a research question to reporting results to the community.

The authors intended the paper as a starting point for computer scientists and domain scientists alike to incorporate neurosymbolic techniques into their research. The same community organized a neurosymbolic programming summer school at Caltech in July 2022 and hands-on workshops on AI for code and science at MIT CSAIL in 2023, which Dr. Costilla Reyes helped organize. A later paper on this site, “Symbolic regression with a learned concept library” (2024), has two of the same coauthors, Atharva Sehgal and Swarat Chaudhuri.

The paper acknowledges National Science Foundation award 1918839, Understanding the World Through Code, as support for the research project, a collaboration among researchers at MIT, Caltech, the University of Texas at Austin, Stanford, and the University of Pennsylvania.