News article by By Moises Velasquez-Manoff.
Published in The New York Times.
Brains are talking to computers, and computers to brains. Are our daydreams safe?
Jack Gallant never set out to create a mind-reading machine. His focus was more prosaic. A computational neuroscientist at the University of California, Berkeley, Dr. Gallant worked for years to improve our understanding of how brains encode information — what regions become active, for example, when a person sees a plane or an apple or a dog — and how that activity represents the object being viewed.
By the late 2000s, scientists could determine what kind of thing a person might be looking at from the way the brain lit up — a human face, say, or a cat. But Dr. Gallant and his colleagues went further. They figured out how to use machine learning to decipher not just the class of thing, but which exact image a subject was viewing. (Which photo of a cat, out of three options, for instance.)
One day, Dr. Gallant and his postdocs got to talking. In the same way that you can turn a speaker into a microphone by hooking it up backward, they wondered if they could reverse engineer the algorithm they’d developed so they could visualize, solely from brain activity, what a person was seeing.
The first phase of the project was to train the AI. For hours, Dr. Gallant and his colleagues showed volunteers in fMRI machines movie clips. By matching patterns of brain activation prompted by the moving images, the AI built a model of how the volunteers’ visual cortex, which parses information from the eyes, worked. Then came the next phase: translation. As they showed the volunteers movie clips, they asked the model what, given everything it now knew about their brains, it thought they might be looking at. [ . . . ]