Researchers at Google-owned DeepMind in the UK have developed AI that can store knowledge, such as a map, and use it to navigate a system as complicated as London's Underground.
Sure, you can already get directions from Google Maps to navigate transport networks, but DeepMind's new system inches it towards the goal of building a neural network that can navigate without any human-written programming, instead using knowledge to work out a route.
Its latest efforts combine deep-learning algorithms with a machine equivalent of a human's working memory.
This so-called 'differential neural computer', or DNC, can use its memory to produce answers from scratch, according to DeepMind. It learns to use that memory through a training process involving comparing answers it produces with the correct answer, which over time get closer to the desired answer.
"We wanted to test DNCs on problems that involved constructing data structures and using those data structures to answer questions. Graph data structures are very important for representing data items that can be arbitrarily connected to form paths and cycles," explained DeepMind researchers Alexander Graves and Greg Wayne.
In a new paper published in Nature, the researchers demonstrate that a DNC can learn on its own to write down a description of an arbitrary graph, such as a map of the London Underground, family trees, or story snippets, and answer questions about them.
"When we described the stations and lines of the London Underground, we could ask a DNC to answer questions like, 'Starting at Bond street, and taking the Central line in a direction one stop, the Circle line in a direction for four stops, and the Jubilee line in a direction for two stops, at what stop do you wind up?'. Or, the DNC could plan routes given questions like 'How do you get from Moorgate to Piccadilly Circus?'," the researchers write.
As noted by IEEE Spectrum, the addition of external memory vastly improved a neural network's ability to answer the second question correctly. Without external memory, a neural network had an average accuracy of 37 percent after two million training examples. The DNC reached an average of 98.9 percent accuracy.
Graves told The Guardian the work was a step towards smart machines, rather than a sudden departure.
"I'm wary of saying now we have a machine that can reason," he said. "We have something that has an improved memory, a different kind of memory that we believe is a necessary component of reasoning. It's hard to draw a line in the sand."