Google DeepMind has created a program able to navigate the London Underground by combining deep learning and a computerised short-term memory.
The London-based company published a paper on Nature detailing the invention of a “differentiable neural computer” (DNC), a system that overcomes the limitations of standard neural networks.
Neural networks are algorithms—modelled after biological neurons—which are able to discover patterns in vast amounts of data. While they are extremely effective when it comes to recognising images, translating languages, or beating world-class players at Go.
On the other hand, these systems are unable to act on data that is fed them only once— for instance, they will not understand the meaning of a text, or making sense of a specific map. That is because standard neural networks cannot store a specific piece of information indefinitely, and after a while they will overwrite it.
To solve that, Google DeepMind combined a standard neural network with an external memory similar to the random-access memory (RAM) found on most personal computers. That in a way harked back to human working memory— a neural warehouse stretching across various brain regions, where recent information is stored.
According to the paper, the new architecture worked. The DNC, given a map of the London Underground, was able to quickly find the shortest route between two stops. While that is not a monumental breakthrough per se— Google Maps or Citymapper have been doing it for years— the mechanism the AI used to carry the task out is innovative, and could likely beget smarter virtual assistants.
The DNC was also able to grasp family relationship after being fed a family tree, and answer simple questions about a short story.
“I’m wary of saying now we have a machine that can reason,”Google DeepMind’s researcher Alex Graves told the Guardian. “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.”
Co-working space and blog dedicated to all things data science.