Brain-scan viewed-image reconstruction study

Reconstructing what volunteers saw

MIT Technology Review reported on October 1 that a Weizmann Institute of Science team developed a system to reconstruct images viewed by volunteers from high-resolution functional MRI scans. According to the report, the researchers trained on existing scan data from eight people who each viewed around 9,000 images. Their decoder estimates image structure and content, while a second model predicts brain activity from images to supplement scarce paired scan data. Researcher Michal Irani said the approach needs about one hour of calibration for a new person, compared with around 40 hours in earlier attempts.

Current limits and privacy questions

The demonstrations involved people viewing images inside an MRI scanner. The system sometimes reconstructed the wrong object, and the team has not shown that it can read spontaneous thoughts, dreams or unwilling subjects. Researchers interviewed by MIT Technology Review raised potential future privacy concerns if related methods become easier to use, but those concerns are not evidence of a present capability to extract anyone’s thoughts remotely. The finding was presented at a conference in September; the October 1 article is the source for these reported details.