Friday, April 01, 2016

The Machine Learning Turk

Google's AlphaGo took the world by storm when it won its match with Lee Sedol but Demis Hassabis now acknowledges the dark truth. Google wanted to promote its cloud computing division as Amazon AWS and Microsoft Azure have quite the head start. Google needed a killer app that would bring users to Google Cloud and decided they could win if they had the best machine learning tools. They bought Deepmind, run by Hassabis, and needed a showcase event and decided to focus on Go, a game yet to be conquered by computers. Hassabis and his team used clever machine learning techniques on top of Monte Carlo Tree Search but only made mild improvements to the game. Google was growing desperate so a plan was hatched.

Using a modern version of the mechanical turk, an 18th century chess playing automaton that secretly hid a human inside playing the game, Hassabis enlisted Japanese Go player Yuta Iyama to secretly choose the moves for AlphaGo. Iyama, who worked with Google when they agreed to remove Iyama's embarrassing Karaoke videos from YouTube, didn't have to physically be in the machine but relayed the moves by a method Hassabis wouldn't reveal. AlphaGo, secretly getting its moves from Iyama, easily dispatched the European champion in October.

Hannabis and his team wrote up their failed algorithms and found it shockingly easy to fool the Nature editors and reviewers. Yann LeCun of Facebook looked at the Google's team's Nature paper and didn't see that much different from what Facebook had tried. "I just figured Google had chosen better parameters to make their program successful. At the time I should have realized what Google was up to."

Google took a risk challenging Lee Sedol but Sedol, not realizing he was really facing Iyama, played the wrong style of game and lost the match four games to one.

Will this revelation hurt the future of AI? "Machine learning continues to change society, but when it comes to Go," said LeCun, "Alpha fools".


  1. Great story for the day. :-)

  2. Any source to this? Can't find anything about it anywhere else...

  3. Thanks for this, Lance. I lectured on adversary search last week in my discrete math course, using xkcd 1002 as a visual aid, so I emailed the students a link to this post to make sure they are fully up with the latest developments in man-machine competition.

    1. One of your students here, is there a way for a computer to play a more complex game with randomness involved. Think of this game like poker stratego and go combined. There are complex decision trees, randomness and hidden information. Or would a computer playing such a game at a competitive level be largely impossible.

  4. "who worked with Google when they agreed to remove Iyama's embarrassing Karaoke videos from YouTube" ...Haha, I can think of someone who might also be solicited by Google soon

  5. Where is the source of this?
    "Hassabis enlisted Japanese Go player Yuta Iyama to secretly choose the moves for AlphaGo."

  6. Just to make it clear, the post above was a complete figment of my imagination. Happy Alpha Fools Day!

  7. You fooled me till "Google was growing desperate so a plan was hatched." :)

  8. Historically one person fell for this: Garry Kasparov seriously thinking Deep Blue was cheating with human assistance in 1997.

  9. This a rather 'in crowd' joke. How many people as a % of the World Population know of this story ? It comes from the age of Casanova. Those famous lines in 'The Third Man' about Switzerland and Cuckoo clocks need to be re-scripted: "Switzerland concentrated on state of the art precision industries aimed at creating decorative and useful machines and instruments". Of course some developers would be tempted to cheat. That's just hype.