Looks like Google has never heard of Skynet. The search giant's secretive X Labs, best known for developing self-driving cars and futuristic eyeglasses, has pioneered an ultra-intelligent machine capable of learning just like the human brain does. As detailed in The New York Times, Google scientists have unleashed the machine on the internet to see what it could learn on its own, and will present their findings at a conference in Scotland this week. Here, a concise guide to the breakthrough:
How does it work?
The machine uses 16,000 computer processors to create roughly one billion connections, mimicking the neural networks used by the brain. Once this network was built, scientists presented the computer with 10 million still images found in YouTube videos to see what kind of information it would teach itself.
What happened next?
Over the course of the experiment, the super-intelligent machine taught itself how to recognize cats. "We never told it during the training, 'This is a cat,'" says project leader Dr. Jeff Dean. "It basically invented the concept of a cat," and burned an image of what the animal is supposed to look like into its mind.
How on Earth did the computer do that?
"The Google brain assembled a dreamlike digital image of a cat by employing a hierarchy of memory locations to successively cull out general features after being exposed to millions of images," says the Times' John Markoff. Essentially, scientists "developed a cybernetic cousin to what takes place in the brains visual cortex." This seems to confirm the idea, says Markoff, that "individual neurons are trained inside the brain" to tell objects apart, which allows a person to differentiate between a face, tree, or shoe.
What did the computer do with its new skillset?
X Lab researchers tested the machine's artificial intelligence by having it comb through some 20,000 pictures of random objects. It was able to successfully pick out images of cats with 70 percent more accuracy than previous efforts. It's a lot like learning to "identify a friend through repetition," says neuroscientist Gary Bradski.
What good is a cat-identifying supercomputer?
"Machine learning is useful for improving translation algorithms and semantic understanding," says Steve Musil at CNET. In other words, the eventual goal is to make Google's diverse line of products more human-like, and thus easier to use.
Sources: All Things D, CNET, Mashable, New York Times