In 1997, IBM’s Deep Blue computer defeated the world chess champion Garry Kasparov. It was a momentous occasion for artificial intelligence, showing how far a clever algorithm could go in a game prized for its intellectual difficulty.
But even as computer scientists celebrated that achievement, there was another game that posed an even greater challenge: the ancient board game Go. Deceptively simple, Go has two players alternately place black and white stones on a 19-by-19 board. But the gameplay is dizzyingly complex, with more possible board positions than there are atoms in the universe.
This is why mastering Go has always had a special appeal for AI researchers: Even the most sophisticated algorithms have fallen far short of matching the best human players.
Until now. Late last year, DeepMind invented a self-learning program called AlphaGo that won 5-0 against the European Go champion Fan Hui. This week in Seoul, South Korea, AlphaGo is taking on the ultimate challenge: a five-match tournament against the world’s best Go player of the last decade, Lee Se-dol. (AlphaGo won the first match.)
This is a remarkable moment for AI — just last year, most researchers and Go experts thought it would take another decade for a computer to take on any professional at Go. Now AlphaGo can rival the top grandmasters.
The AI research community has made incredible progress in five years. Researchers at Google have turned what used to be theoretical machine learning constructs into tangible, useful products. Some of the effects are subtle improvements: more relevant search results, tighter spam filters, better recommendations for videos and apps.
But machine learning also lets you do things people have never been able to do before. Google Photos allows you to ask your phone to find that picture of your daughter in 2013 with the basset hound in the snow, and it appears within seconds. And if you hold up a phone to a sign in Russian, you can get an instant English translation overlaid with Google Translate. At Google we’re starting to teach computers automatically to recognize parts of speech, understand the meaning of sentences and recognize handwriting, images and videos.
Overall, a key insight has been that it’s much better to let computers figure out how to accomplish goals and improve through experience, rather than handcrafting instructions for every individual task. That’s also the secret to AlphaGo’s success.
We first trained AlphaGo’s neural nets on 30 million moves from human games so it could learn to predict what move a human expert would make. Then we went a step further, playing AlphaGo against itself in thousands of games. Doing so allowed it to discover entirely new strategies by learning from its own mistakes, so it could play even better than the people from which it had first learned.
Does that mean that if AlphaGo wins, we lose? Are we diminished when a computer can defeat humans at a game that seems to typify human ingenuity?
Quite the contrary. The real challenges in the world are not “human versus machine,” but humans and whatever tools we can muster versus the intractable and complex problems that surround us. Whether climate change, health care or education, the most important struggles already have thousands of brilliant and dedicated people making progress on issues that affect every one of us.
Technologies such as AI will enhance our ability to respond to these pressing global challenges by providing powerful tools to aid experts make faster breakthroughs. We need machine learning to help us tame complexity, predict the unpredictable and support us as we achieve the previously impossible.
As our tools get smarter and more versatile, it’s incumbent upon us to start thinking much more ambitiously and creatively about solutions to society’s toughest global challenges. We need to reject the notion that some problems are just intractable. We can aim higher.
None of us knows who will be the victor in this week’s matches, and that’s part of the thrill. But away from the excitement of the head-to-head contest, consider how unstoppable the next generation of Go players would be with AlphaGo on their side — and consider what the world’s best climate change researchers, clinicians or educators could achieve with machine learning tools assisting them. The real test isn’t whether a machine can defeat a human, but what problems humans and machines can overcome together.
