Brain Reading Device Can Predict What People See

March 14, 2008

"Brain Reading" Device Can Predict What People See

http://news.nationalgeographic.com/news/2008/03/080305-brain-scan.html

Ker Thanfor National Geographic News

March 5, 2008

A new computer program can match brain activity with visual images and even predict what people are seeing, a study has shown. The work raises the possibility that one day computers could "read" a person's brain to digitally re-create memories, dreams, or imaginings.

Enlarge Photo

Printer Friendly

Email to a Friend

Supercomputing Project Aims to Simulate Human Brain (July 20, 2005)

"Beyond the Brain" in National Geographic Magazine (March 2005)

Previous attempts to decode vision in this way could only extract simple information about images, such as their physical orientation, and could not identify images that participants were seeing for the first time. "Our technique overcomes this limitation, and we show that we can perform identification for novel images," said study team member Kendrick Kay of the University of California, Berkeley. The new computer model is described in today's issue of the journal Nature. Master Predictor The researchers used functional magnetic resonance imaging (fMRI) to measure activity in the visual cortices of participants' brains as they looked at photographs of animals, food, people, and other common objects. The fMRI technique is a relatively new way to measure changes in the brain's blood oxygen levels, which have strong links to neural activity. The collected data were used to "teach" a computer program to associate certain blood flow patterns with particular kinds of images. Participants were then asked to look at a second set of images they had never encountered before. The model was programmed to take what it had learned from the previous pairings and figure out what was being shown in the new set of images. For a collection of 120 images, the model correctly identified what a person was looking at 90 percent of the time. When the set was enlarged to a thousand images, accuracy was about 80 percent.