AI decodes monkey visual neurons into human language with 96% accuracy

Understanding what individual neurons encode has long been a core challenge in neuroscience, especially for higher visual areas where simple mathematical models like Gabor filters fall short. Researchers from Stanford, Baylor College of Medicine, and other institutions have now demonstrated that natural language can fill this gap. Using digital twins of macaque primary visual cortex (V1) and area V4, they developed a closed-loop framework that first translates each neuron's high- and low-activating images into dense captions, then generates a semantic hypothesis and synthetic images to test it. The system runs entirely in silico, enabling rapid, scalable experimentation on neural representations.

The results are striking. In V4, activating hypotheses drove 96.1% of neurons above the 95th percentile of natural-image responses, while suppressing hypotheses drove 97.6% below the 5th percentile—compared to roughly 10% for random images. V1 activation results were similarly strong, though suppression in V1 was less amenable to semantic description. Representational similarity analysis showed that vision embeddings align most closely with neural activity, and while the text encoding step loses some information, that alignment is recovered when hypotheses are rendered back into images. This reveals that linguistic compression is lossy yet semantically faithful. The work points toward AI-driven scientific discovery, where agents can automatically generate testable hypotheses about neural function.