Skip to main content
Event

AI for Code and Science Workshop

On February 2–3, 2023, Dr. Omar Costilla Reyes co-organized the MIT AI for Code and Science Workshop, where he also co-led a session. The session covered neurosymbolic programming for molecular generation.

Background and introductory remarks for the AI for Code and Science Workshop. Watch on YouTube.
Announcement graphic for the two-day workshop, February 2–3, 2023; the owner’s post credits the illustration to OpenAI DALL-E. Credit: Image courtesy of Dr. Omar Costilla Reyes. Illustration generated with OpenAI DALL-E.
A February 2, 2023 event-day photograph shows the MIT CSAIL schedule listing Dr. Costilla Reyes and Jose Cambronero as speakers for AI for Code and Science Day 1. Credit: Image courtesy of Dr. Omar Costilla Reyes.
A workshop session at MIT, February 2023. Credit: Image courtesy of Dr. Omar Costilla Reyes.
Dinner after the first day of the workshop, February 2023. Credit: Image courtesy of Dr. Omar Costilla Reyes.

A hands-on workshop at MIT

The MIT AI for Code and Science Workshop ran on February 2–3, 2023, as a hybrid event at MIT, announced as part of the Institute’s Independent Activities Period (IAP) 2023. Dr. Omar Costilla Reyes announced it as the first hands-on workshop of the Neurosymbolic Programming for Science project at MIT CSAIL and organized it in collaboration with Jose Cambronero of Microsoft Research; the CSAIL schedule for the first day, February 2, lists both of them as speakers. The aim was practical experience: participants with some programming background were to work with AI tools to develop code for scientific applications, with time to experiment with working code and to get feedback on their own ideas.

AI for code and science

AI for code and science is an active research area that connects machine learning, programming languages, and software engineering, and much recent progress in it blends symbolic and neural techniques. Neural networks learn patterns from data but are hard to read from the inside; symbolic methods, such as program synthesis (the automatic generation of programs from examples or specifications), state their steps explicitly and can be checked. The workshop’s program covered how to combine the two and apply them in science, with hands-on use of transformer-based code models such as CodeBERT, CodeT5, and Codex. Dr. Costilla Reyes gave the background and introductory remarks on neurosymbolic programming for science, and, with Minghao Guo of MIT CSAIL, led a session that applied the approach to molecular generation, using a learnable graph grammar that builds molecules from production rules and can be trained on small chemical datasets.

The other sessions were Miles Cranmer of Princeton on symbolic regression with PySR, which searches for simple mathematical expressions that fit data and can translate a neural network into an interpretable form; Jose Cambronero on using language models to repair compiler errors in C programs; and Shashank Srikant of MIT CSAIL on how the brain responds to reading code and how a program can be modified to change that response.

Recordings and what followed

Dr. Costilla Reyes shared the first recording on February 6, 2023 and, on March 24, the recordings and code of the full workshop through the Neurosymbolic Programming for Science project, noting that they range from code repair to applications in physics, chemistry, and neuroscience. The materials make the sessions and demonstrations available beyond the original event. The workshop belongs to the same research line as the Neurosymbolic Programming Summer School, whose first edition took place at Caltech in July 2022, and he later gave a webinar on AI for code and science at Universidad Peruana Cayetano Heredia; each has its own page on this site.

Topics