Subtitles section Play video Print subtitles So...computers get really good at beating us at chess. And then they get really good at beating us at a lot of things. We're so used to getting beat by computers that when I play Greece in deity mode, I don't cry any more when it stomps me. 'Kay, when I play Greece in any mode, I don't cry when it stomps me. But humanity, crafty devils that we are has long had one secret weapon...Go! And it was just a matter of time before those computers came for that too. Welcome to CompChomp...the only show on the internets where we can talk about atari without having to mention any video game crashes. From the beginning of AI work, GO was considered the impossible dream. There were games like tic-tac-toe that could be completely solved with math. And if you can solve it with math, computers are gonna dominate. There were games like chess that were a bit harder, but with a little bit of tweaking of the algorithms, you could also make computers dominate there too. Hey...uhh...didn't I just talk about this? Go is really different though. You play it by taking stones and placing it on a 19 by 19 grid. And the goal is to surround more territory than your opponent does. There are more possible game combinations than there are atoms in the entire known universe. Possibly the unknown universe too. We haven't counted it. And every single piece has the exact same value. So that chess approach...it ain't gonna work. As recently as 2015, experts in AI were predicting that it would be at least...oh...a decade before any computers were able to beat top Go players. Then, in January of 2016, the journal Nature came out with a little article talking about the algorithms that the Google-backed AI had used to defeat the European champion. And, BTWs, that exact same AI was going to be facing off against one of the top players in the world, Lee Sedol, in March. Mic drop. So what had changed between those expert predictions of at least a decade and that article in Nature? Advancements! Advancements in Machine Learning. Ahhhhhh...Machine Learning. It's my favorite! SInce Go has so many possible moves during any game, that uh whole more game combinations than atoms in the universe thing, you can't have the computer go through every possible move to determine which move is the best. Instead, computers use an algorithm called the Monte Carlo Tree Search. This is where they take a random sampling of possible moves, calculate out to see which one of those is the best, and then select that move. AlphaGo is no different than any of the other Go-playing computers on this front. What sets AlphaGo apart is it uses Deep Learning to prune the unwanted combinations before it does the Monte Carlo Tree Search. What's Deep Learning? So glad you asked! Deep Learning is a kind of artificial neural network inspired by the fancy little brains of mammals. Brains have neuron and artificial neural networks have simple processors. Those neurons in brains can be connected. And so can those processors. Brains have a way of getting input from the outside world. Things like your eyes and your ears. And artificial neural networks have that too. There are processors that can take in data from the world around it. Similar to how the brain passes messages from neuron to neuron in order to make decisions, data gets passed through the layers of an artificial neural network with each layer transforming it in different ways based on what it's "learned" in previous layers - until decisions are made. So AlphaGo's training started out with an initial data set of 30 million training moves. It worked with this data until it could correctly identify the next human move 57% of the time. So, just a little over half of the time, it could look at what a human was playing and say, "I know what you're gonna do!". Then, they matched it up with a slightly modified version of itself and it just played thousands of game. AlphaGo vs AlphaGo Prime! Who will come out on top?!? And, after that, the European champion played hundreds of games against it so it got used to playing against humans too. After alllllll of these games, AlphaGo was ready. This face off was amazing. Game 1 starts and AlphaGo and Lee Sedol trade moves. Back and forth. Back and forth. For over 3 hours! They were mostly testing each other. Trying to get a sense for how the other one played. And at the end, even though he had led the entire game, Lee Sedol lost to AlphaGo. Computers - 1. Humans - 0. The second game got off to a very similar start. They were tentative. Back and forth. Back...and forth. Then, on move 37, AlphaGo's piece gets plunked down....*pop*...on a random part of the board. It's called a shoulder hit, and everyone there that knows anything about Go was just like, Whaaaaaaaaaaaaaaaaa! The people that built AlphaGo thought it made a mistake. They looked at the control room data afterwards and this was not even one of the moves that was in the 30 million training moves. I mean, think of that. 30 million moves by top players, and, not a single one of them was this move that AlphaGo made. Lee Sedol stared at the board. And then he gets up. And he walks out of the room. And he's gone for 20 minutes. So he comes back in. And he sits down. And he makes his move. Move 38. Move 39. But you can tell he's shaken. He just never brings it back. And this game goes on for hours. But then he resigns. And now....it's AlphaGo - 2 Humans - 0. But do not be afraid! Don't worry. Because this...this is THE moment that humanity rises up and we throw off the yoke of our new computer overlords. No....this isn't some Hollywood movie. This is real life. And this is a top AI system. Humans lost. We lost 4 of the 5 games. So we're not completely defeated, but, Go is no longer a human-only realm. AlphaGo had made his mark. Or her mark. Or its mark. So with this loss, should we run for the hills and hide from our robot overlords? No! I mean, AlphaGo knows how to play Go. Ask it to play chess and it'd probably do worse than my kid sister. Well, I haven't got a kid sister. Chomp!
B1 alphago ai deep learning chess move neural Deep Learning and AlphaGo - Computers Beating Humans #2 143 7 Caurora posted on 2017/01/27 More Share Save Report Video vocabulary