Eric James Beyer in Quanta:
First, he and his colleagues test people on tasks related to language or vision. Then they compare the observed behavior or brain activity to results from AI models built to do the same things. Finally, they use the data to fine-tune their models to create increasingly humanlike AI.
The process works best with more data and more models, so Schrimpf built an open-source platform called Brain-Score(opens a new tab) that contains nearly a hundred human neural and behavioral data sets. Researchers have tested thousands of AI models against the human data since Schrimpf first developed the platform in 2017, back when he was still in graduate school.
Schrimpf originally planned to work in the tech industry, but after co-founding a pair of software startups during his early academic career, he felt unfulfilled. “I thought I could ask neuroscientists how the brain works, and that would help me build better AI,” he said. “But I realized there’s a huge opportunity in the opposite direction: prototyping ideas in silico [on a computer] and using AI models to explain the brain.”
More here.
Enjoying the content on 3QD? Help keep us going by donating now.