DQN plays Atari
A network learned to play seven Atari games from screen pixels and the score alone, and beat a human at three of them.
Why it matters
One algorithm with no knowledge of the rules learned different tasks from scratch, and that is what convinced the field reinforcement learning scales.
Earlier systems needed features written for each game. DQN received what a person receives: an image and a number. The main technical devices were an experience replay buffer and a separate target network, which stabilised training. Google acquired DeepMind within six months of publication. The 2015 Nature version covered forty-nine games.