Go Go Gadget AI: Why the AlphaGo Victories Matter
This past week, artificial intelligence enthusiasts around the world have been abuzz over a Chinese board game called Go. If you’re not familiar with it, it’s a game of strategy that’s been captivating people for thousands of years. What’s putting it in the headlines these days, however, is the fact that a computer program called AlphaGo (brought to you by the folks at Google DeepMind) is currently beating one of the best Go players in the world. As I write this, AlphaGo is beating Lee Sedol three games to one, with the fifth game set to go off later today.
If your first response to all of is along the lines of “so what,” let me explain why it’s actually a really big deal.
While the goal of the two-player game is simple — using stones to try to surround more territory on the board than your competitor — it’s incredibly complex to play. In fact, until recently Go was widely believed to be too complex for computers to be able to play well.
All of that complexity stems from the sheer number of configurations possible for placing the black and white stones on the board. There are 2.08168199382×10170 possibilities, to be precise, when using a standard board with a 19×19 grid of intersecting lines. That’s more configurations than there are believed to be atoms in the universe.
What’s interesting about AlphaGo’s approach is that rather than trying to model all the possible moves completely in advance, it’s teaching itself to play the game using a technique called reinforcement learning. To become proficient, the software has played thousands of games with itself and other players to discover new strategies through a process of trial and error.
This video from DeepMind (made before the ongoing match with Lee Sedol) provides a quick overview of the complexity involved in playing Go. You can also read more in this blog post from the DeepMind team. If that’s not enough to satisfy your curiosity, the folks at Nature published this excellent video that serves as an introduction to the game and the approach AlphaGo uses. (You can get even more of the nerdy details in this paper Nature published describing the approach.)
And if you like podcasts as much as I do for keeping up with everything that’s going on in the technology industry, then here are a couple that cover the recent AlphaGo victories and why they are so significant:
The a16z team do a good job of putting the magnitude of AlphaGo’s victories in context and, in the process, provide a useful taxonomy of artificial intelligence approaches. Perhaps more important are their points about the power of AlphaGo’s “ensemble” approach, whereby the team used multiple machine learning techniques in combination rather than adopting a purist approach to deep learning. They also make the point that that kind of pragmatic approach to AI innovation is the way that product designers should be tackling AI challenges.
AI Safety and the Legacy of Bletchley Park (Talking Machines)
Nick Patterson, of the Broad Institute, talks to the Talking Machines team about why he is so surprised and excited by AlphaGo and its ability to play Go well. He also gives some good insights into the data-driven process behind AlphaGo vs. the more rules and brute force approach used in the past to conquer the much simpler game of chess. Nick also points out that Google has not published how much compute time it took to train the software; he hypothesises that it could have been up to a million hours to get to this point.
I will keep this list updated as more content comes out. It’s a big breakthrough and worth understanding given it’s likely to lead to further opportunity and disruption across the software industry and beyond. If you’ve seen or heard something great on the subject that we’ve missed let us know in the comments below.