One of the major milestones in the relatively short history of AI is when Google’s AlphaGo beat the best human Go player in the world in three straight games early last year. In order to prepare AlphaGo for its match, Google trained it using games played by other Go players, so it could observe and learn which moves win and which don’t. It learned from essentially watching others.
This week, Google announced AlphaGo Zero, AI that completely taught itself to win at Go. All Google gave it was the rules, and by experimenting with moves on its own, it learned how to play, and beat its predecessor AlphaGo 100 games to zero after just over a month of training.
Why It’s Hot:
AI is becoming truly generative with what DeepMind calls “tabula rasa learning”. While a lot of AI we still see on a daily basis is extremely primitive in comparison, the future of AI is a machine’s ability to create things with basic information and a question. And ultimately, learning on its own can lead to better results. As researchers put it, “Even when reliable data sets are available, they may impose a ceiling on the performance of systems trained in this manner…By contrast, reinforcement learning systems are trained from their own experience, in principle allowing them to exceed human capabilities, and to operate in domains where human expertise is lacking.”