For those of you who have never heard of it (I certainly hadn’t), Go is an ancient Eastern game involving both strategy and intuition. Unlike chess, Scrabble, Othello and Jeopardy, Go has baffled AI engineers for decades, allowing Europe’s reigning Go champion, Fan Hui, to keep his privileged position as the best Go player in the world.
Unfortunately for Hui, his reign has now ended. Researchers at DeepMind, a collective affiliated with Google that has called itself an “Apollo program for AI”, developed AlphaGo, an AI system that beat Hui five games in a row during a machine-versus-man contest in October.
“It was one of the most exciting moments in my career, both as a researcher and as an editor,” said Dr. Tanguy Chouard, an editor for Nature and witness of the event.
AlphaGo uses DeepMind’s deep learning system. Deep learning is an increasingly important aspect of AI technology that allows for computers to sort information on their own when given a huge amount of information. Without deep learning, a computer could only ever be as good as the Go moves recorded into its system. With deep learning, it’s able to understand why those moves are good and then generate a new collection of moves that could top any human Go master.
“The most significant aspect of all this… is that AlphaGo isn’t just an expert system, built with handcrafted rules,” explained Demis Cassabas, who oversees DeepMind. “Instead it uses general machine-learning techniques hot to win at Go.”
Deep learning is used by a variety of service providers to create technology that makes it possible for computers to identify images, recognize spoken words, and understand natural language. DeepMind actually combines deep learning with something called reinforcement learning as well as the methods and claims that this combination of learning programs is “a natural fit for robotics.” Soon robots are likely to learn how to perform physical tasks and respond to their environment in ways that will make them helpful to society, or make them more effective killers in times of war.
Cassabas believes that these methods can also help to accelerate the advancements of scientific research:
“The system could process much larger volumes of data and surface the structural insight to the human expert in a way that is much more efficient- or maybe not possible for the human expert,” he stated. “The system could even suggest a way forward that might point the human expert to a breakthrough.”
For now, researchers are still testing the reach of AlphaGo. Its next competitor will be Lee Sedol, who holds more international titles than Hui and has the second-most in the world.
Cassabas believes that the ball is still in AlphaGo’s favor.
“Go is implicit. It’s all pattern matching,” Hassabis explained. “But that’s what deep learning does very well.”
Only time shall tell on what the limits of AlphaGo are, or the technological limits of deep learning and the research at DeepMind, for that matter. Hopefully killer robots don’t become too much of a thing.