What it is
Researchers recorded large volumes of neural activity from the visual cortex of many mice and trained a transformer-based foundation model to predict how individual neurons respond to arbitrary natural videos. The model generalized to entirely new mice with minimal additional data and predicted responses to stimulus classes it was never trained on, such as coherent motion and noise patterns. Beyond activity, it also predicted anatomical cell types, dendritic features, and neuronal connectivity in the MICrONS connectomics dataset.
Why it matters
Earlier brain-activity models tended to break down outside their training distribution, which limited their scientific use. Showing that a single pretrained model captures transferable latent structure of neural computation brings the foundation-model paradigm into systems neuroscience. The result has already drawn 72 citations, signaling rapid uptake as a template for brain models.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed underfoundation model, neuroscience, visual cortex, connectomics, generalization