At first glance, the victory of Google’s AlphaGo AI over South Korean Go grandmaster Lee Sedol looks like it belongs in the same category as IBM’s triumphs — first with Deep Blue over Garry Kasparov and later with Watson over Jeopardy! celebrity Ken Jennings. However, while it’s not incorrect to group these examples of artificial intelligence‘s advancement together, it might be more apt to refer to AlphaGo as a sudden leap forward, whereas its predecessors were merely laying a foundation.
And just what is AlphaGo leaping toward? If its Go performance is taken at face value, the answer could be a faster and more intelligent Internet of Things.
First of all, it warrants mentioning the challenge that a game like Go presents to a computer that chess and “Jeopardy!” — for all their intellectual airs — do not. While the latter two test a player’s ability to react to an opponent’s move and recall specific information, respectively, Go is game simply of controlling spaces on a board. There are no specialized pieces, but the American Go Association (AGA) explains that each game contains the possibility of a staggering number of moves — 1 duocennovemnonagintillion, or 10 to the power of 900, to be “exact.”
The AGA politely notes that that represents more subatomic particles than we think exist in the universe.
But crunching big numbers is exactly what supercomputers and AIs are supposed to do, right? That would be correct if modern computer science focused on creating nothing but giant calculators. However, Harvard Business Review outlined the two major factors that drove AlphaGo to four victories and only one loss over Sedol, despite his apparent skill and experience: a policy network and a value network. Though the names might sound a little more mundane than what they do, HBR summarized the first as helping AlphaGo determine the moves that would best guarantee an overall victory, while the second identified the best possible actions related to the individual move at hand.
When these two networks of information interact, they helped AlphaGo not only plot its course to ultimate victory but also ensure that it maintained a sufficient lead over Sedol. In an interview with Maclean’s, artificial neural network expert Geoffrey Hinton explained that this ability to simulate “intuition” through data analysis — AlphaGo had access to thousands and thousands of past Go games — could mean big things for IoT and how the machines around us might stop responding to taps and clicks and start listening to language instead.
“It’s already influencing things. In Gmail, you have Smart Reply that figures out from an email what might be a quick reply, and it gives you alternatives when it thinks they’re appropriate,” Hinton said. “They’ve done a pretty good job. You might expect it to be a big table, of ‘If the email looks like this, this is a good reply, and if the email looks like that, then [t]his might be a good reply.’ It actually synthesizes the reply from the email. The neural net goes through the words in the email and gets some internal state in its neurons and then uses that internal state to generate a reply. It’s been trained in a lot of data, where it was told what the kinds of replies are, but it’s actually generating a reply, and it’s much closer to how people do language.”
Hinton took issue with calling AlphaGo a true AI, as it didn’t quite learn how to play, as much as it checked possible moves against the likelihood that they’ve worked in other games and would work in the current one. However, Google was able to develop a program that beat an 18-time Go world champion years before most pundits said it would ever be possible. That can only bode well for the exponential progress AlphaGo and other intuitive software will make in the half-decade they now have free.