CMU Researchers Are Teaching Computers To Read Minds

Jul 3, 2017

Carnegie Mellon University researchers have taught a computer to “read minds,” so to speak.

Using data from functional MRI scans, the computer was able to identify sentences such as “the man kicked the football” and “the tourist went hiking in the forest” with 87 percent accuracy.

Cognitive neuroscience professor Marcel Just said the team scanned human subjects’ brains while they read and thought about 240 different sentences. Different parts of the brain were activated depending on the words used in the sentences.

“Such as whether they pertain to a person, an emotion, a physical action,” Just said. “If it’s an object, maybe its size or its color. If it’s an activity, maybe it’s a mental activity or a physical activity. If it’s a physical activity, maybe it involves your feet more or your hands more.”

Just’s team then taught the computer which areas of the brain are associated with which concepts.

To test the computer’s learning, they fed it the brain activation patterns that corresponded with 239 sentences. Then, they gave it the activation pattern for a left-out sentence, the only one the computer hadn’t already learned.  

“It goes back and look and says ‘Which of these 240 sentences does this match up with best?’ And usually it does very, very well in guessing which one,” Just said.

He said it’s not exact at this point. For example, the computer can’t distinguish between “man” and “person,” but it can distinguish between “person” and “car” or “punching” and “kicking.”

Just said he imagines creating an archive of brain activation patterns for all kinds of different words and concepts. He said such an archive is possible because most people’s brains activate in the same way when they think about particular words, even among people who speak different languages.

“You can train the program up on a bunch of other people, and then a new person comes in and you can tell what they’re thinking,” Just said.

He said the technology could be useful in diagnosing psychiatric disorders, such as schizophrenia or obsessive compulsive disorder.

“A psychiatric illness is a distortion of thought. It means something about thinking has been distorted,” Just said. “With this approach, you can measure whether certain kinds of thoughts are distorted.”

Just said previous research had shown that a similar approach is highly accurate in identifying people with autism, by scanning their brains while they consider social concepts such as “insult” and “persuade.”

He said there are also possible applications in better understanding the learning process and how the brain organizes information, so that instructional methods can be better tailored to the brain’s organizational system.

The study is published in the journal Human Brain Mapping.