Insight

Man vs Machine: What AlphaGo’s Triumph Tells Us

The epic battle between Google DeepMind’s AlphaGo and Go grandmaster Lee Se-dol has kept many people at the edge of their seats. It marked a historical milestone for artificial intelligence (AI) when AlphaGo sealed a 4-1 victory over Lee, which leads to the big question: will AI surpass and even replace humans soon?

In the same week as AI’s success, IAS invited Prof Qiang Yang, New Bright Professor of Engineering and Chair Professor of Computer Science and Engineering at HKUST, to share his knowledge of and insights into AI and AlphaGo in an IAS Commons session. Here are some of the more exciting areas covered in the talk.

IAS

Why is AlphaGo a significant step forward for AI?
 

Qiang

In 1997, IBM’s Deep Blue beat the best human chess player at that time. Go, however, is a game of profound complexity. There are 1047 possible games in chess. But with 150 moves in each game, Go has 10700 possible games—even more than the total number of atoms in the universe.

What also makes AlphaGo special is that it teaches itself to play. DeepMind first trained AlphaGo by feeding it 160,000 actual gameplays between humans, then the team allowed it to play against itself to generate 30 million extra samples and master the game through reinforcement learning.

IAS

AlphaGo only defeated the 2-dan European champion last November. How did it improve so much in five months to be able to crush a 9-dan grandmaster?
 

Qiang

AlphaGo was designed to be just a little better than its opponent. Its primary objective is to win. Even winning by one point is good enough for AI, unlike humans who tend to maximize their success. When AlphaGo was pitted against the European champion, it only performed slightly better than a 2-dan player. Playing against stronger opponents made it stronger as well.

\\  In the future, people who have the best job security are likely to be programmers and those at the top of their fields.  \\

IAS

Why did AlphaGo lose the fourth game?
 

Qiang

Lee made a brilliant decision in the 78th move, which other human players described as “the hand of God.” Yet AlphaGo evaluated it as an amateur move—a mistake that no human would make. AlphaGo was trained through data input and by playing against itself. Whether the opponent was Lee or itself made no difference to AlphaGo. Ultimately, its lack of awareness of the reality led to its downfall in the fourth game.

IAS

Will AlphaGo or other AI systems replace humans soon?
 

Qiang

There is still a long way to go before computers can do what we can. Take games for example—even though AI has managed to beat top human players in chess and Go, it has only tied with human poker champions so far. Chess and Go are games with perfect information, meaning that all of the pieces are out in the open. Poker is a game of imperfect information. To win, one must also be able to see through opponents’ bluff, which is no mean feat for computers. Unless computers can set their objectives on their own, they will not completely replace humans. Whoever manages the objectives will therefore control the AI systems and what they do. In the future, people who have the best job security are likely to be programmers and those at the top of their fields.

IAS

Where is AI going next?
 

Qiang

AI systems like AlphaGo are designed for specific purposes. If taken outside of their pre-determined domains, they will fail miserably. What developers like DeepMind want to achieve next is generality in AI, which means developing systems that will excel not just in Go or chess, but also in other areas such as financial markets, online education and recommendation systems.

IAS

How are HKUST’s pursuits in AI development?
 

Qiang

HKUST launched a joint AI lab with WeChat last November. More than 10 professors and students are developing AI systems that can do a variety of things: tell stories in photos and videos, recognize human speech, process emotions, and undergo transferred learning, the last of which can hopefully fix the reality bias in AlphaGo’s fourth game. Another interesting project is an AI that can predict the financial market, which is already doing quite well in its investment decisions.