Mechanistically Interpretable Neural Encoding Reveals Fine-Grained Functional Selectivity in Human Visual Cortex

In three linesMINE (Mechanistically Interpretable Neural Encoding) applies mechanistic interpretability tools to neural encoding models to localize visual features driving voxel-level activity in human visual cortex. Validated via image generation and counterfactual editing: inserting/removing predicted features shifts neural activation as expected.

Summary generated by Claude — human-verified