If you’ve been reading this blog for any extended period of time (and for that, we thank you), you know we get pretty excited about developments in Google’s AlphaGo program. While the two-player strategy game of Go might not be the most practical application of machine learning and artificial intelligence, it provides an amazing use case for testing a neural network’s ability to learn, adapt and grow at an incredibly complex task.
Most of what we’ll ask A.I. to do in the future will only be more complex than Go, not less. As such, it stands to reason that Google spending time and effort teaching its A.I. behemoth to master the most complex strategy game on earth is time, money and manpower well spent. But what happens when humans have taught the machines all we can? What then?
Well, we’re starting to get a glimpse of it now.
At its heart, a neural network or deep learning consists of a supercomputer being fed mountains of data and then using precisely honed algorithms, the neural network picks out patterns, trends and the like from the data, and teaches itself how to act accordingly. The more complete the data, the more data available, the more the machine can learn and progress.
But, this method of machine learning has a massive drawback — it requires the machine to think like a human.
That concept is great when you’re talking about talking; we want our machines to emulate humans when it comes to voice recognition and vocal responses. But forcing a machine to think like a human when it comes to complex strategy? Not so much.
Humans are great at pattern recognition — in fact, it may be what we’re best at. But we’re also prone to biases, which can be fatal in strategy games. So, once AlphaGo had gotten to a point where it was decimating the top players in the world, Google made a radical tack and removed humans from the equation entirely.
The result? It beat the previous version of itself — the same version that spanked the top-rated human player in the world — in a landslide. A 100-point landslide, that is:
In a paper published in Nature earlier this week, DeepMind revealed that a new version of AlphaGo (which they christened AlphaGo Zero) picked up Go from scratch, without studying any human games at all. AlphaGo Zero took a mere three days to reach the point where it was pitted against an older version of itself and won 100 games to zero — Wired.
That’s a pretty staggering margin of victory, considering the loser was the best player the world had ever seen. What’s almost more interesting, though, is how human players are responding to the games played. The new incarnation of AlphaGo, dubbed ‘Zero,’ hasn’t played any public games yet. But several months ago, Google released 55 games an older version of AlphaGo played against itself. Since May, human experts have spent countless hours analyzing those games, and their descriptions seem to focus on the same sentiments: “Amazing. Strange. Alien.” per Wired.
“They’re how I imagine games from far in the future,” Shi Yue, a top Go player from China, has told the press. A Go enthusiast named Jonathan Hop who’s been reviewing the games on YouTube calls the AlphaGo-versus-AlphaGo face-offs “Go from an alternate dimension.” From all accounts, one gets the sense that an alien civilization has dropped a cryptic guidebook in our midst: a manual that’s brilliant—or at least, the parts of it we can understand.
It will be fascinating to see how the progress of AlphaGo Zero proceeds. But interestingly enough, it’s no longer the machine learning from human data, but rather humans learning from machine data.
What a brave new world, indeed.