夢に出てきた物の種類を、脳の活動から当てる。京都の研究チームが挑んだ「夢の解読」

Dreams are very private. Usually, only the dreamer knows what they saw. In 2013, a team in Kyoto, Japan, tried to read dreams from brain activity.

Three people slept in an MRI machine. They were woken up more than 200 times each and asked about their dreams. A computer program learned their brain patterns. Then it could guess things in their dreams, like cars or people, better than chance.

Dreams have always been difficult to study because scientists could only ask people what they remembered. In 2013, a team led by Kamitani in Kyoto tried a new approach. Three participants slept inside an MRI scanner. Each time brain waves showed they were falling asleep, they were woken and asked what they had seen. This was repeated more than 200 times per person.

Next, the researchers showed the participants many photos while they were awake and trained a machine-learning program to recognize objects from their brain activity. When the program was given brain activity from the dreaming period, it predicted the kinds of objects in the dreams better than chance. This suggests that seeing and dreaming use similar brain patterns.

In 2013, three people slept in an MRI scanner at ATR in Kyoto. Whenever EEG showed that one of them was drifting into sleep, a phase rich in dream-like images, they were woken and asked what they had just seen.

Kamitani's team wanted to decode dream content from brain activity. Dreams have long resisted objective study, since only the dreamer can report them. The awakenings were repeated more than 200 times per person, and the reports were sorted into categories such as cars, people and buildings.

Next, the researchers trained decoders on brain activity recorded while the participants were awake and looking at photos. They then applied the decoders to activity measured just before each awakening. The decoders identified the categories of dreamed objects at rates above chance, and higher visual areas were especially informative.

So dreaming and seeing appear to share similar neural patterns. The method has clear limits. It needs extensive training data from each person, it picks out only broad categories, and it studied images at sleep onset instead of the story-like dreams of REM sleep. Nobody could use it to read a stranger's dreams. The team later decoded imagined objects and reconstructed viewed images.