MINE Framework Decodes Voxel-Level Visual Features in Human Cortex

A new framework called Mechanistically Interpretable Neural Encoding (MINE) opens the black box of how artificial neural networks predict brain activity. Developed by researchers and published on arXiv (2605.16468), MINE applies mechanistic interpretability tools to localize the specific image features driving millimeter-scale (voxel-level) responses in human visual cortex. Unlike prior correlational approaches, MINE uses language-aligned image representations to produce semantically interpretable descriptions of critical features for each voxel's activation. The method generalizes per-image features to reveal fine-grained functional selectivity, advancing understanding of what visual content drives neuronal activity.