These days, artificial intelligence (AI) is capable of a lot of different things. Machines can come up with new beers, write films, mimic sounds, detect cancer, and beat humans at incredibly complicated games. And now, a new AI has figured out how to ride the subway.
In a new study in Nature, artificial intelligence researchers at Google’s DeepMind report that they have developed a machine learning model they call a differentiable neural computer. It combines a computers’ ability to process complicated data with the machine learning ability of neural networks. This system has both memory, which allows it to make inferences and recall facts, and the ability to learn from previous experiences.
They proved its capabilities, among other experiments, by having it tackle a task that can be hard even for people: It had to plan out the best route in an unfamiliar subway system—in this case, the London Underground.
Because of its memory capabilities, it can reason its way through such a map system using what it has learned about other maps. In this case, the AI system was first trained on a series of graphs where it had to traverse certain routes and find the shortest route between points. After completing training of 1 million examples, it achieved 98.8 percent accuracy on these types of problems.
Unlike previous systems, this one doesn’t need to be manually programmed. Instead, it can be trained to complete a task through examples or through trial and error. It can learn to understand relationships in graphs like subway maps or family trees, and find common connections.
DeepMind hopes to further develop this type of artificial intelligence to tackle more complicated machine-learning tasks, like language processing or cognitive mapping.