Artificial intelligence has come a long way in the past few decades, and recently it’s taken a step forward that has left some folks feeling a little dismayed. A computer program is now able to beat the best Go players in the world—something it hasn’t been able to do before. And it did so by learning. The computer was able to study millions of examples of the game and learn how to beat human competitors. NHPR's Peter Biello spoke with David Brooks, a reporter for The Concord Monitor and writer at Granitegeek.org, about "deep learning."
I don’t mean to make light of the potential dangers of artificial intelligence, but people like Stephen Hawking have said that the fact of computers being able to learn quote “could spell the end of the human race.”
Yeah, that’s kind of sobering terminology. He’s not the only extremely intelligent person to put out similar cautions recently. Elon Musk is the other high-profile one we may have heard, basically saying artificial intelligence (and, to a certain extent, robotics) is making strides that are kind of alarming as well as exciting. And as you say, I talked about the game of Go, an ancient East Asian game, as an example.
Let’s step back and talk about the game. It originated in China hundreds of years ago. And it’s more complicated than chess, which is why computer programs have been able to play chess with some success, but not Go—at least, not until recently.
“Complicated” is a tough word. I like to call it “deeper than chess,” not necessarily harder. So, as you say, it originated in East Asia, and a lot of people play it in the states, but it certainly is dominated by Korean and Chinese players.
For context, can you tell us about the object of the game?
It’s a board game. It’s a 19 by 19 board, as compared to chess, which is an eight by eight board. And you take turns putting colored stones, either white stones or black stones down, and you try to surround more territory than the other person.
A year ago, really, there were no computer programs that did well against professional players, whereas of course computers can beat the best at chess now. So it seemed like it was something that you had to be human to really understand.
This computer program that’s able to beat human players of Go—it’s using something called “deep learning.”
Deep learning. So, AlphaGo is the name of the program. It’s a Google project designed specifically to tackle Go. Deep learning is a terminology for a way of learning. A lot of chess programs—Deep Blue from IBM— examine eighty gazillion options for each move and decide which one is the best option. What “deep learning” does is study games that have been played and use the result to change the weighting of its own program, which is exactly the way our brain works. As you learn, as you grow, it really changes the strength and connection between neurons in your brain and that’s really what memories are and that’s what our personality is, to a certain extent. And that’s very, very analogous to what deep learning does. It’s an approach to teaching a program.
The success this has had—because, as I say, a year ago, computer Go programs were semi-laughable, and last month they beat Lee Sedol. I don’t know if he’s the best Go player in the world, but he’s up there. And it didn’t just beat him. It crushed him, 4-1, including one move in the second game, 37th move, that Sedol had to leave the room for 15 minutes, even though it’s a timed game, simply because it’s a startlingly brilliant move. It was amazing. People were still talking about it.
So that’s led people like me to say, “Woah, maybe they really are kind of going too far. Maybe carbon-based life forms are in trouble.”
So computers can, as science fiction predicted, learn from experience and grow. Is Go the final frontier for this kind of technology? Something tells me it isn’t.
Of course not. And some people argue that people like me are going way overboard, that Go is this limited board game and it’s mathematical, so it’s the sort of thing a computer could tackle. So quite possibly I’m freaking out unnecessarily. But I think it’s an example of how quickly this method of approaching a problem, how quickly it allows artificial intelligence to tackle problems that seemed insurmountable. It makes you wonder if it’ll turn to solving other things, like the best way to live life, how government should be run, how to do your job well.