The trick, however, is that Facebook is carefully tracking everything the human operators do, so it can feed this information into neural networks and teach machines to perform the same tasks. That is, they visit the restaurant's website and make the reservation. Today, while under test with a few hundred users in the San Francisco Bay Area, M is largely driven by human operators: you ask the tool a question along the lines of "Can you make me a dinner reservation for tonight?," a fairly simple AI system suggests at least a partial solution, and then the humans perform the task. Meanwhile, the company aims to build an even higher form of AI through the digital assistant it calls M. "You take reasoning-the ability to ask questions and understand new data-and you take image understanding and segmentation and you put the two together," Schrep explained, "and you build what we call visual Q&A." Image "segmentation" is where the system correctly distinguishes between different objects in a photo-separating, say, the baby from the man. "So, we've taken some of the basics of game-playing AI and attached a visual system to it, so that we're using the patterns on the board-a visual rec system-to tune the possible moves the system can make." Though this system is only about two or three months old, he says, it can already beat systems built solely with more traditional AI techniques. "We're pretty sure the best players end up looking at visual patterns, looking at the visuals of the board to help them understand what are good and bad configurations in an intuitive way," Facebook CTO Mike "Schrep" Schroepfer told reporters at Facebook's California headquarters last week, before delivering a speech along similar lines this morning at the Web Summit in Dublin. Researchers are feeding images of Go moves into a deep learning neural network so that it can learn what a successful move looks like. Now, Facebook is using similar technology to recognize a promising Go move-to visually understand whether it will be successful, kind of like a human would. And Microsoft can instantly translate your Skype calls.
Google's smartphone digital assistant can recognize the commands you bark into your Android phone. Thanks to these neural networks, your Facebook app can recognize photos of you and your friends.
To recognize a cat, for instance, a deep learning system analyzes thousands of known cat photos, feeding each into a network of machines that approximate the neural networks of the human brain. In recent years, companies like Facebook, Google, and Microsoft have shown that deep learning is remarkably adept at recognizing photos, identifying spoken words, and translating from one language to another. With this in mind, researchers at Facebook are now tackling Go with an increasingly important form of artificial intelligence known as deep learning. You can't use the same approach as a Deep Blue or a Watson. Getting a computer to play this way is another task entirely. The top players will tell you they play in a way that's, on some level, subconscious. On a Go board-a 19-by-19 grid where players place pieces at the intersection of two lines-the number of possible moves is far greater, and identifying the benefits of a particular move is far more complicated, even mysterious. Likewise, a machine can look ahead in a game of Go- the Eastern version of chess-but in this case, looking ahead is far more difficult. But a machine can examine far more future moves than Kasparov ever could. Yes, a chess grandmaster like Kasparov can look ahead in a similar way. With all those other games, computers can win by, in essence, analyzing the many possible outcomes of every possible move. But there's one notable pastime where we humans still come out on top: the game of Go. Machines can now beat the best humans at a wide range of games traditionally held up as tests of intellect, from Scrabble to Othello.
And in 2011, another IBM machine, Watson, topped the best humans at Jeopardy!,the venerable TV trivia game show. Three years later, to much fanfare, IBM's Deep Blue supercomputer won its chess match against reigning world champion Gary Kasparov. In the mid-'90s, a computer program called Chinook beat the world's top player at the game of checkers.