Scientists Are Trying to See Our Dreams
Who can capture dreams? Researchers try, making videos from brain scans and dreamscapes from machine learning.
Google’s network of artificial neurons generated this image of its “dreams.” Image (computer generated) by Michael Tyka, Google
Scientists are still far from understanding how humans generate dreams, but Google engineers have tapped into the dreamlike “imagination” of a computer. These images were created by letting a system turn a randomly patterned background into images it recognizes.Do androids, as sci-fi novelist Philip K. Dick asked, really dream of electric sheep? The purpose and meaning of dreams have long been debated. Now scientists are getting closer to deciphering what humans see as they sleep—and how a robot can simulate it.
In 2013 neuroscientist Yukiyasu Kamitani had test subjects take hundreds of brief naps in an MRI machine, repeatedly waking them so they could describe their dreams. Kamitani had already isolated the unique brain patterns for certain objects he’d shown subjects while awake. Their brains were scanned for those patterns as they napped, and a computer program automatically turned the basic contents of their dreams into short videos. The study found these were 70 percent accurate compared with what subjects remembered of their real dreams.
Two years later Google engineers also captured the dreamlike images of a computer. They fed millions of images into a brain-inspired computer algorithm—a network of artificial neurons—to study how it learned to identify objects. Then they put it through DeepDream, a program that enables the network to build its own algorithm-fueled dreamscape by finding shapes in an image of random visual noise, like the static on an old TV. The computer generated a psychedelic scene from its machine-learned knowledge. As in a dream, previously seen images were reconfigured into new patterns.
It won’t be possible to produce a precise recording of human dreams until scientists discover how dreams originate in the brain, says Jack Gallant, a professor of psychology at UC Berkeley—or they build an encyclopedia of brain activity that corresponds to every thought. He likens it to building a language translation program: “You have a language but nothing it references to.”
An artificial neural network—a computer algorithm inspired by the human brain’s natural network—was first fed millions of images to teach it to identify objects on its own. For example, to learn what a dog looks like, it was shown images of all types of dogs.
After training a computer to identify certain objects, software engineers at Google reversed the system, giving it a prompt and asking it to produce its own.
The engineers then put the computer through a program called DeepDream. They provided it with a canvas of random noise, like the black-and-white static on an old TV, and let it pull out shapes it could identify.
Artificial neural networks are programmed to find patterns in abstract things, just like people see images in clouds. In DeepDream it perceives these shapes and then renders them increasingly recognizable.
After repeating this process multiple times, landscapes, figures—like the birds seen here—and objects emerged. The network was producing images as simple as geometric patterns and as complex as full buildings.
Though the system can’t be compared with the human imagination, it could perhaps be used to understand the creative process—or at the very least, inspire some creativity.