For the first time, a computer has beat a person in the board game Go. The program was developed by Google - reason enough for Mark Zuckerberg, who's developing his own Go program, to formulate a new AI goal.Katharina Schneider
Mark Zuckerberg could celebrate his empire's new record numbers, but had to swallow a competitor's success at something he would rather have won himself: a program developed by Google company Deepmind beat a human player for the first time in the Asian board game of Go.
Facebook is also working on Go software, but with less success than Google. Reason enough for Zuckerberg to set himself a new goal in the field of artificial intelligence (AI): a system that can assist him in the office and at home - like Iron Man's Jarvis. And he wants to teach it common sense.
This might sound like something out of science fiction or a Marvel comic, but AI and learning software are already commonplace, as evidenced by Facebook's newsfeed (which learns which types of post users most often click on), Amazon's product recommendations, the Spotify stream that gets to know listeners' musical tastes better over time, and Google search results. All these are based on algorithms that are fed with user behavior data, where past behavior is used to predict what users will want in the future. This is known as predictive data analysis.
Developing machines that can think much as humans do cannot only mean kudos for the creators, but also plenty of money. So it's no wonder that the digital giants are battling it out for superiority in the field of machine learning.
Google Now and Apple's Siri, the speech recognition systems on Android and iPhone, are already practically at the level of personal digital assistants. Siri does what the user tells her to. But that's not enough. In the future Siri should be able to do what Google Now already can: anticipate what the user will want next.
Apple is investing heavily to keep their competition in Mountain View from irrecoverably eclipsing them. Last fall the company was looking for 86 new employees purely for AI development. Then in January it acquired startup Emotient, which has developed technology that uses AI to recognize feelings by analyzing facial expressions and associated emotions. Advertising agencies are using this software to analyze the reactions of campaign audiences. Even doctors have used Emotient, with patients who have difficulty communicating.
It's not clear what Apple is aiming for, but the tech giant has a big problem ahead: AI systems learn from data. And Tim Cook recently declared that – unlike the competition – Apple respects what the Germans refer to as "informational self-determination", and will not collect user data. After all, as Cook points out: Apple earns its money primarily with hardware.
This is dipolar to Google's approach – an approach that has them labeled as a data vampire. More than five years ago, Google's CEO Eric Schmidt was already quoted as saying: "We know where you are. We know where you've been. We can more or less know what you're thinking about." The logical continuation of this statement is: We know what you want, even before you know it yourself. Anticipating behavior is the goal of every AI. But you can only know what behavior to expect when there's lots of data available. Where does the user normally spend time? Where does he or she shop? What is his day like? Who are his or her closest contacts?
Now the company could reach a new milestone in AI. The "AlphaGo" AI, developed by Google acquisition Deepmind, was able to beat professional Go player Fan Hui. The Asian board game of Go is considered to be even more complex than chess. Until now, humans have been able to beat every computer. Go has always been viewed as the final bastion of human intelligence, and therefore as a yardstick for AI developers. Whoever can get their computer to win the game is the leader in the field of artificial intelligence.
Although this game was played last fall, the success of Deepmind's Software is just now being recognized with the publication of an article by David Silver and Deepmind founder Demis Hassabis in Nature. The authors celebrate this victory as "a feat people thought was still decades away". In March, AlphaGo will compete against the player considered the best in the world, South Korean Lee Sedol.
Google managed this exploit, not only with huge amounts of data, excellent developers and early investment in this field, but also by relying on the intelligence of the masses rather than secretive development in isolated rooms. In November the company made the Tensorflow program package available as open source software. According to Sundar Pichai, Tensorflow is faster, smarter and more flexible than older AI systems – and largely responsible for Google's progress in the field of artificial intelligence.
By making this "secret weapon" available free of charge, Pichai hopes to speed up development by means of mass efforts. "We hope that the AI research community will be able to share ideas much faster in the future - in the form of functional program code instead of theoretical scientific papers," says Pichai.
Facebook also developed AI that can learn Go last year. Zuckerberg reports that Facebook is not far from completing an operational Go program, and wastes no words on Google's success.
But shortly thereafter Zuckerberg's gauntlet came down: he wants to develop a Jarvis AI. First it will learn from patterns. This is a little bit like showing picture books to a little boy or girl and asking them what they see, explains Zuckerberg. Any problems that can be reduced to pattern recognition can be solved by artificial intelligence in this way.
To start with, the Facebook founder says he's satisfied that "his" Jarvis recognizes him at the door and automatically opens it for him, or lets him control his house with his voice. But his AI system will not only recognize patterns; it will also have common sense, because this is how knowledge can be taken from one area and applied to another. This allows machines to react to problems and situations it doesn't know. The only problem? "No one knows how to teach it to AI."
Zuckerberg thinks that this can only happen through "unmonitored learning." Returning to the child comparison, this would be like handing the child a book and letting him or her figure out what to do with it. He sees this as the key to machines with human-like common sense – an independently thinking robot that can talk and plan complex actions.
People shouldn't be afraid of AI, according to Zuckerberg. The biggest unsolved problem of our times is to find out how learning works. Until we have a firm grip on that, no computer can ever become as smart as a person.