Dream contents deciphered by computer

View the video

A computer can decode the stuff of dreams. By comparing brain activity during sleep with activity patterns collected while study participants looked at certain objects, a computer learned to identify some contents of people’s unconscious reveries.

“It’s striking work,” says cognitive psychologist Frank Tong of Vanderbilt University in Nashville, who was not involved in the research. “It’s a demonstration that brain activity during dreaming is very similar to activity during wakefulness.” 

The work, reported April 4 in Science by Japanese researchers led by Yukiyasu Kamitani of Advanced Telecommunications Research Institute International, adds to somewhat scant knowledge of how the brain constructs dreams, says Tong. The research could lead to a better understanding of what the brain does during different states of consciousness, such as those experienced by some coma patients.

Dreams are a bit of a black box and difficult to study. Experiments with mice have revealed aspects of sleep and dreaming, such as how the experiences contribute to forming memories. But a mouse can’t tell you what it dreamed about. And the sleep stage that’s richest in dreams — REM sleep — typically kicks in about 90 minutes after a person conks out, making it time consuming to gather data on dreams. The noisy fMRI brain scanning machine doesn’t help.

To skirt these experimental issues, the researchers recorded brain activity in three adult male volunteers during the early stages of sleep. After the subjects had dozed off, they were repeatedly awakened and asked for detailed reports on what they had seen while sleeping. In an example, one participant stated: “Well, there were persons, about three persons, inside some sort of hall. There was a male, a female and maybe like a child. Ah, it was like a boy, a girl and a mother. I don’t think that there was any color.”

After gathering at least 200 such reports from the three men, the researchers used a lexical database to group the dreamed objects in coarse categories, such as street, furniture and girl. Then the study participants looked at images of things in those categories, while their brains were again scanned. Computer algorithms sorted through these patterns of brain activity, linking particular patterns with objects.

When the computer went back to the brain scans taken during dreaming, it did a pretty good job of distinguishing some of the signals, such as whether a dream contained a book or a girl. On average, the computer could pick which of two objects had appeared in a dream 70 percent of the time, a rate that is much better than would be expected by chance.

“To be able to get enough data to do this kind of analysis is really impressive,” says Russell Poldrack, a neuroimaging expert at the University of Texas at Austin.

The study bolsters the notion that the vivid imagery of dreams, no matter how fantastic, is as real as waking life, Kamitani says, at least from the brain’s perspective. Further research may reveal if the same is true about other dreamed senses, such as experienced sounds or emotions.

After monitoring a man’s brain activity while he looked various kinds of objects (top images), scientists trained a computer to decode patterns of brain activity recorded while the man dreamed. The computer could distinguish among many dreamed objects. The size of the word in the word cloud (bottom) indicates the likelihood that that object was in the dream.
Credit: Courtesy of Yukiyasu Kamitani, ATR