Researchers Compress Vision Models To Predict Cortex

In a Nature paper, researchers led by Benjamin Cowley trained large DNNs on macaque V4 responses and compressed a 60-million-parameter model to roughly 1/1,000 its original size. The compact model predicted neural activity more than 30% better than prior state-of-the-art and revealed interpretable units, including V4 "dot" detectors. The approach enables testable circuit hypotheses and could inform visual-stimulation therapies for synapse loss.