Skip to content
BENIBIX Learn

Deep Blue to AlphaGo

Level 4 · AI history

Three moments that changed AI. A chess win by search in 1997, a photo contest in 2012, and a win at Go by a program that learned.

Course contentsLesson 9 of 10

After this lesson you can

  • Explain how Deep Blue won at chess
  • Say why the 2012 ImageNet result mattered
  • Tell how AlphaGo learned in a new way

12 mincovers 3 glossary entries

Worth reading first: Dartmouth and the AI Winters

Deep Blue, ImageNet and AlphaGo: what each one proved

Deep Blue, a chess computer built by IBM, beat world champion Garry Kasparov in May 1997, by 3.5 games to 2.5. It checked about 200 million board positions every second. It did not learn chess the way modern AI does. It won by searching ahead, with scoring rules that chess experts helped write.

Kasparov had won their first match in 1996, by 4 games to 2. The rematch made news around the world. Yet after the win, many said it was just fast search, not real thinking. That shrug, once a machine can do a thing, comes back in almost every story of AI.

ImageNet is a huge set of labelled photos, about 14 million of them, built so computers could learn to see. From 2010, a yearly contest asked teams to name what was in its photos, across 1,000 kinds of thing. For two years, the best programs still got about one photo in four wrong.

In 2012, a team from Toronto entered a network with many layers, trained on graphics chips. On the contest's main score it got 15.3% wrong, against 26.2% for the next best entry. A jump that big in one year had never happened. Labs switched to learning from huge piles of data, and the deep learning boom began.

AlphaGo, from DeepMind, plays Go, a board game with far more possible moves than chess. Searching ahead like Deep Blue was hopeless. AlphaGo first learned from games played by people. Then it played millions of games against itself, getting better each time. Nobody wrote its plan of play by hand.

In March 2016, in Seoul, AlphaGo beat Lee Sedol, one of the best players in the world, by 4 games to 1. One of its moves in game two looked like a mistake to experts and turned out to be brilliant. A later version that learned only by playing itself beat it 100 games to 0.

Together the three show the shift from rules to learning. Deep Blue followed rules that people helped write. AlphaGo found moves no person had taught it. But people often read these wins as general smarts. Each system was narrow, and could do nothing else at all.

The early wins followed rules or search. The later ones learned from data.

Worked example

Emma counts the moves

Emma plays both chess and Go. On a typical turn, a chess player has about 35 moves to pick from, and a Go player has about 250.

Looking four turns ahead means about 1.5 million lines in chess and about 4 billion in Go. That gap is why Go needed a program that learned, not just one that searched.

The words for this lesson3

Check yourself

How did AlphaGo get good at Go?

Finished this lesson?

Marking it complete also marks its 3 glossary entries learned, which updates your roadmap and your saved page. 0 are already marked.