EP.000 · Research talk · NeurIPS 2022
Why Interpretable? Code as the Language of Science
Dr. Omar Costilla Reyes · Host of Interpretable, founder of equ healthcare
Hosted by Dr. Omar Costilla Reyes
Listen or watch
Audio edition, 32 min. Keyboard: K play/pause, J back 15 s, L forward 30 s. The audio edition skips the tutorial logistics, the two-day agenda, and the group discussion; they are in the video.
Full video, 43 min.
Chapters(17)
About this episode
Deep learning finds patterns in data, but scientists need models they can read, question, and build on. In this talk, Omar Costilla Reyes introduces neurosymbolic programming: models that combine neural networks with symbolic programs written in a domain-specific language. He explains what a program is, how program synthesis works, and why writing a model as code makes it interpretable, because much of scientific knowledge already takes the form of code: equations, rules, and procedures. The talk opened a two-day tutorial on AI for code and science in the Neurosymbolic Programming for Science series, and it explains the idea behind this show's name.
Listen if: You want the big picture behind interpretable AI for science and health, and the idea that gives this show its name.
The research behind the episode
NeurIPS 2022 Workshop on AI for Science2022Peer reviewed
Neurosymbolic Programming for Science
Jennifer J. Sun, Megan Tjandrasuwita, Atharva Sehgal, Armando Solar-Lezama, et al.
In plain language
The paper argues that neurosymbolic programming, which pairs neural networks with symbolic programs, can speed up scientific discovery while keeping results readable. Because part of each model is a program, scientists can build their own domain knowledge into it and inspect what it learned.
Why it matters for health
Health decisions need models people can check. Writing models as programs is one route to AI that clinicians and scientists can read, question, and trust.
Key findings
- 01Neurosymbolic models can take scientists' symbolic domain knowledge as input and produce interpretable results.
- 02Behavior analysis serves as a worked example of where the methods help today.
- 03The authors map the opportunities and open challenges for the natural and social sciences.
What you'll learn
- What program synthesis is, and how it differs from training a neural network
- Why writing a model as code makes it interpretable to scientists
- How neurosymbolic programming connects machine learning with scientific knowledge
Working on something like this?
Suggest a guest or a paper for a future episode, or talk to equ about AI for health in your own work.
Keep listening
Your Brain on Code: What Brains and AI Models Both Learn
Shashank Srikant
NeurIPS · 2022Paper discussed
Using brain recordings to read what AI models of code have learned, and what people represent when they read programs.
The Simplified Mental Maps We Plan With
Mark K. Ho
Nature · 2022Paper discussed
A precise, testable model of how people simplify problems before they plan.