Researchers reconstruct viewed images from high-resolution brain scans
A reported fMRI system produced closer reconstructions of pictures seen by volunteers, while still making visible mistakes and requiring scanner data.
A narrower result than mind reading
Researchers at Israel’s Weizmann Institute of Science have built a system that reconstructs images people looked at while their brain activity was recorded in a high-resolution functional MRI scanner, according to an October 1 report by MIT Technology Review. The research team used existing data from eight volunteers who had each viewed around 9,000 images. One part of the system predicts the layout and colors of a viewed image, while another predicts its content; an image-generation model then combines those estimates into a reconstruction.
Researcher Michal Irani told the publication that a second model, which predicts brain activity from an image, helped the team use additional images without corresponding human scans for training. She said the resulting decoder needs about an hour of calibration for a new participant, versus roughly 40 hours in earlier approaches. Those comparisons are the researcher’s account in the report, not an independent replication presented here. The work was shown at a conference in September, and the reviewed article does not establish how the method performs outside the described research setting.
Where it still fails
The reported examples include clear errors: a cake was reconstructed as sandwiches and a dog in a bathtub as a goat. Irani hopes future versions may reconstruct imagined scenes or dreams, but she explicitly said that capability does not exist yet. Experts interviewed by MIT Technology Review raised concerns about mental privacy as brain-decoding methods improve. This experiment required cooperative participants and an MRI scanner; it does not show that thoughts can be read at a distance or without consent.