Brain-IT AI Reconstructs Images from Brain Activity in One Hour

The Weizmann Institute of Science's Prof. Michal Irani has introduced a groundbreaking AI model named Brain-IT, capable of accurately reconstructing images from brain activity in roughly one hour—much quicker than earlier models that took several hours. This system utilized over 70,000 images presented to eight subjects during brain scans. To address the limited dataset, researchers developed an encoder to recognize common brain activity patterns among individuals. Each scan was segmented into about 40,000 voxels, allowing AI to identify visual elements like colors, locations, faces, and food. The team pinpointed 128 functional regions linked to image processing, revealing both familiar and novel areas, including distinct roles within the parahippocampal place area (PPA). MIT Technology Review reported that the encoder and decoder collaboration enhances image reconstruction accuracy, with potential uses in aiding communication for those with paralysis and advancing the understanding of visual processing in the brain.