EP.005 · Research talk · NeurIPS 2022
Your Brain on Code: What Brains and AI Models Both Learn
Shashank Srikant · Runs Manana Labs, a boutique consulting group specializing in AI and deep tech
Hosted by Dr. Omar Costilla Reyes
Listen or watch
Audio edition, 38 min. Keyboard: K play/pause, J back 15 s, L forward 30 s. The audio edition covers the first 38 minutes, the neuroscience part of the session. The hands-on coding tutorial that follows is in the video.
Full video, 2 hr 13 min.
Chapters(10)
About this episode
What happens in the brain when someone reads a program? In this session, Shashank Srikant reviews cognitive neuroscience research that uses fMRI recordings of programmers reading Python. Two brain systems that support very different kinds of thinking, the Multiple Demand system and the language system, each encode specific properties of code. He then shows how machine-learning models of code can be aligned with these brain recordings, and how programs can be modified so that a model predicts stronger responses in particular brain regions. That comparison is an interpretability method in its own right: lining up a model's internal representations with brain signals shows what the model encodes, and where it differs from people. It is why this episode sits on Interpretable.
Listen if: You are curious about cognitive neuroscience, brain imaging, or methods for understanding what AI models learn.
The research behind the episode
NeurIPS 20222022Peer reviewed
Convergent Representations of Computer Programs in Human and Artificial Neural Networks
Shashank Srikant, Benjamin Lipkin, Anna Ivanova, Evelina Fedorenko, et al.
Advances in Neural Information Processing Systems 35: 18834–18849
In plain language
The study uses fMRI recordings of programmers reading Python to ask which properties of a program the brain represents. It scores static and runtime properties of code against brain signals, then maps brain recordings onto the representations learned by machine-learning models of code.
Why it matters for health
Comparing brains with AI models is a growing tool in cognitive neuroscience: it shows what each system has learned, and where models of the mind need to improve.
Key findings
- 01The Multiple Demand system and the language system both encode specific properties of code.
- 02Both systems align with the representations learned by machine-learning models of code.
- 03The results point to at least two distinct neural mechanisms behind understanding and evaluating programs.
What you'll learn
- How neuroscientists use fMRI to study reading and understanding code
- Which brain systems are involved in program comprehension, and how they differ from reading language
- How AI models of code can be compared with brain activity, and used to probe it
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
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.
Designing Molecules With Rules You Can Read
Minghao Guo
ICLR · 2022Paper discussed
A generative model for chemistry built from learned rules: valid by design and readable by chemists.