Researchers at Cold Spring Harbor Laboratory, Carnegie Mellon, and Princeton have compressed an AI vision model from 60 million variables to just 10,000 — a reduction to one-thousandth of its original size — while maintaining nearly the same performance. The study, published in Nature on March 3, used macaque monkey brain data to mimic how V4 neurons process colors, textures, and curves.

"This is something we could send in a tweet or an email," said lead author Ben Cowley, assistant professor at Cold Spring Harbor Laboratory. The compact model reveals how individual neurons specialize — some responding to curved produce shapes, others to eye-like dots — suggesting biology-inspired architectures could lead to far more efficient AI systems.

Mitya Chklovskii, a group leader at the Simons Foundation's Flatiron Institute, called the work significant for understanding both biological and artificial intelligence. The findings suggest current AI models may be vastly over-parameterized and that brain-inspired compression could make powerful AI accessible on everyday devices.