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.
Topic
The science of how people perceive, remember, reason, and plan, and the computational models that make those processes precise.
2 episodes
Cognitive science studies how people perceive, remember, reason, and plan, and builds computational models that make those processes precise. Two episodes of this podcast take it up. In Episode 004, Mark K. Ho presents value-guided construal, an account from a 2022 Nature paper in which people simplify a task into a mental representation that is easier to plan with. His maze-navigation experiments suggest that what people notice and remember depends on what is useful for the plan.
In Episode 005, Shashank Srikant looks at what happens in the brain when someone reads a program. The NeurIPS 2022 paper behind the episode uses fMRI recordings of programmers reading Python and reports that the Multiple Demand system and the language system each encode specific properties of code. It also finds that both align with representations learned by machine learning models of code. In both, a model is set beside measurements from people: behavior in the maze experiments, and brain activity in the fMRI study. For the show this is also a way to understand AI, because lining up a model's representations with human measurements shows what the model has learned. The episodes and papers are listed below.
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.
Mark K. Ho
Nature · 2022Paper discussed
A precise, testable model of how people simplify problems before they plan.