Name: Two Minute Papers - How DeepMind Conquered Go With Deep Learning (AlphaGo)
Uploaded: 2016-09-08T15:09:43.000Z
Duration: 3 min 15 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

Dear Fellow Scholars, this is Two Minute Papers with Károly Zsolnai-Fehér.

In 1997, the news took the world by storm - Garry Kasparov, world champion and grandmaster chess player

Basic passive voice

was defeated by an artificial intelligence program by the name Deep Blue.

In 2011, IBM Watson won first place in the famous American Quiz Show, Jeopardy.

In 2014, Google DeepMind created an algorithm that that mastered a number of Atari games by working on raw pixel input.

This algorithm learned in a similar way as a human would.

This time around, Google DeepMind embarked on a journey to write an algorithm that plays Go.

Go is an ancient chinese board game where the opposing players try to capture each other's stones on the board.

Behind the veil of this deceptively simple ruleset, lies an enormous layer of depth and complexity.

As scientists like to say, the search space of this problem is significantly larger than that of chess.

So large, that one often has to rely on human intuition to find a suitable next move,

therefore it is not surprising that playing Go on a high level is, or maybe was widely believed to be intractable for machines.

This chart shows the skill level of previous artificial intelligence programs.

The green bar is shows the skill level of a professional player used as a reference.

The red bars mean that these older techniques required a significant starting advantage to be able to contend with human opponents.

As you can see, DeepMind's new program's skill level is well beyond most professional players.

An elite pro  player and European champion Fan Hui was challenged to play AlphaGo, Google DeepMind's newest invention

and got defeated in all five matches they played together.

During these games, each turn it took approximately 2 seconds for the algorithm to come up with the next move.

An interesting detail is that these strange black bars show confidence intervals,

which means that the smaller they are, the more confident one can be in the validity of the measurements.

As one can see, these confidence intervals are much shorter for the artificial intelligence programs than the human player,

likely because one can fire up a machine and let it play a million games,

and get a great estimation of its skill level, while the human player can only play a very limited number of matches.

There is still a lot left to be excited for, in March, the algorithm will play a world champion.

The rate of improvement in artificial intelligence research is accelerating at a staggering pace.

The only question that remains is not if something is possible, but when it will become possible.

I wake up every day excited to read the newest breakthroughs in the field, and of course,

trying to add some leaves to the tree of knowledge with my own projects.

I feel privileged to be alive in such an amazing time.

As always, there's lots of references in the description box, make sure to check them out.

Thanks for watching and for your generous support, and I'll see you next time!