Google pips Facebook to solve ‘grand challenge’ of AI (Wired UK)

Peter Dahlgren via Flickr / Creative Commons Licence

DeepMind, Google’s London-based AI business, claims to have solved one of the “grand challenges” of artificial intelligence.

The company, which was purchased by Google for £400 million in July 2014, said it had beaten a professional human player at the ancient Chinese board game Go. The AI system DeepMind created, dubbed AlphaGo, was unbeaten — winning five games in a row — against the game’s European champion Fan Hui. It was also almost undefeated against other computer systems playing Go, winning 499 times out of 500.

“This is the first time that a computer Go program has defeated a human professional player, without handicap, in the full game of Go – a feat that was previously believed to be at least a decade away,” explained the DeepMind research paper — Mastering the game of Go with deep neural networks and tree search — published in Nature.

AlphaGo combined an advanced tree search with deep neural networks, which worked in tandem to process the board through 12 different network layers. One element of the network selects the AI’s next move while the other simultaneously predicts the ultimate winner of the game. The neural networks then played games between themselves to learn new strategies by trial-and-error.

The announcement from Google comes hours after Facebook CEO Mark Zuckerberg grabbed headlines by claiming the social network was “getting close” to solving same problem.

It isn’t clear if Facebook’s algorithms were tested against professional human competitors, nor is the timing of its announcement likely to be accidental — DeepMind had already announced it was working to solve the problem. In December, Facebook AI researcher Yuandong Tian told WIRED US there was a “friendly rivalry” between the two firms, adding there would be “pride” for the company who cracked Go first.

AlphaGo’s matches against Fan Hui

DeepMind via Nature research paper

The game, which is played by 40 million people around the world, involves people taking turns to place black or white stones on a board in a bid to capture an opponent’s stones and territory.

Go has proved to be a significant challenge for AI researchers over the last 20 years. With more positions than chess, estimates have put the number of ways the game could be played at 10 to the power of 700.

“We trained the neural networks on 30 million moves from games played by human experts, until it could predict the human move 57 percent of the time (the previous record before AlphaGo was 44 percent),” explained DeepMind co-founder Demis Hassabis. “We are thrilled to have mastered Go and thus achieved one of the grand challenges of AI.”

Hassabis said AlphaGo would now face Lee Sedol, the world’s top Go player, to further test its credentials.

If the article suppose to have a video or a photo gallery and it does not appear on your screen, please Click Here

27 January 2016 | 6:00 pm – Source:


Leave a Reply

Your email address will not be published.