Decoding the Brain with AI with Taylor Beck
"We won't see a human-like AI until we make one that can lose its mind."
What students do
Over five weeks, students build working versions of brain imaging analyses using public datasets and Python libraries. They start by training a model to figure out what a person was looking at based purely on patterns of brain activity, reproducing one of the foundational results in the field. They move to studying how the brain organizes itself even when nothing is being asked of it. They finish with the analysis that turns out to be the most direct conceptual bridge to AI alignment research: comparing how similar things look across brain regions, individual people, and species, monkey versus human for example, much the way researchers now compare how similar things look across layers of a language model.
The capstone, optional but worth aiming for, is a direct comparison between how a brain represents categories and how a small language model represents the same concepts: what differences, if any, require a body made of cells? Students leave with a working notebook, a written report, and the ability to read papers in this field and know what they are looking at.
About the mentor
Taylor Beck first encountered the tools we now call AI twenty years ago, decoding brain activity from fMRI in a neuroscience lab at Princeton. In labs from Kyoto to Washington University in St. Louis he studied memory, sleep, and aging. This summer his students work with real fMRI datasets to investigate how the brain and a language model represent the same things and what the differences reveal about both.