Dreams defy even the dreamer, slipping away as stealthily as they arrive in a mind made credulous by sleep. But what if scientists could read our dreams by using the most advanced medical imaging machines and employing the sophisticated algorithms that flag fraudulent transactions among millions of credit-card purchases?
Researchers in Japan have taken an early step toward this chimerical goal by training computers to recognize the images flitting through the minds of sleepers in the earliest stages of dreaming. Their results, published online by the journal Science, suggest that machines may be able to read our minds at least while were in the anteroom of dreamland.
Were all intrinsically interested in dreaming, but neuroscientists to this day arent certain what it does for us, said Jack Gallant, who studies the brains visual system at University of California at Berkeley. It would be great to have a method of decoding to allow us to know what is going on when we dream.
Like other researchers applying their brains to the brain, the Japanese scientists are probing dreams to understand how they relate to such core functions as memory consolidation and learning.
The researchers put three volunteers into a functional magnetic resonance imaging machine capable of tracking blood flow in the brain, a sign of neurons at work. They also hooked up the volunteers to electroencephalograph machines, which record the electrical activity of those neurons.
Then the scientists waited for the subjects to fall asleep.
The EEG readings showed when the volunteers entered an early stage of dreaming called hypnagogic hallucination. The researchers woke the subjects roughly 200 times, about every six minutes, to get verbal reports of what they saw before the images faded from memory.
The volunteers responses were understandably groggy, such as: I saw a person ... it was something like a scene that I hid a key in a place between a chair and a bed and someone took it. The researchers focused on the nouns in these descriptions and combined them into generic categories, which were represented by images a human face, a key, furniture and presented to the subjects while awake.
The rest was a giant math problem. The scientists wrote a computer program to sort through the patterns of brain activity captured by the functional MRI in both waking and sleeping states; then the program looked for links between those brain activity patterns and specific images.
The computers learned to decode dream imagery with an average accuracy of 60 percent, according to the study. In some cases, the accuracy was significantly higher.
For some categories like male, female and other characters you can predict if this character was in the dream or not with an accuracy of 70 percent to 75 percent, said study leader Yukiyasu Kamitani, a neuroscientist at ATR Computational Neuroscience Laboratories in Kyoto.
One of Gallants computer programs managed to identify 92 percent of images presented to waking subjects using only MRI readouts. Other studies were able to discern a brains attempt at motor control, which could lead to help for people who have lost control of their limbs.