51����

Memory: upgraded. DeepMind’s latest AI has a “working memory” so that it can learn how to solve tasks for itself – such as how best to get from A to B on the London tube network.

“The thing can learn to compute what it has to, rather than being programmed,” says Murray Shanahan at Imperial College, London, who wasn’t involved with the work.

Called a Differentiable Neural Computer (DNC), the system succeeds because it combines neural networks, which are good at learning but not so good at storing data, with an external memory. It can retrieve items from its memory in the order they were recorded –  a key innovation that ensures they don’t get overwritten too quickly and helps the system tackle complicated data it hasn’t seen before.

The DNC works out how to interpret a data set on its own, following some basic training on random graphs. What’s more, it intuitively learns how to use its working memory appropriately when faced with a task.

One of the tests set for the DNC involved finding the shortest path between two stations on the London tube map. To do this, it had to remember the various connections between individual stations in order to compare one route with another – the point where a traditional neural net stumbles, because it doesn’t have the capacity to remember those details while working.

Quickest journey

The DNC successfully identified a number of journeys with the fewest stops. One route it suggested – Moorgate to Piccadilly Circus via Bank and Holborn – is probably a few minutes slower than going via King’s Cross, a journey suggested by apps like Citymapper and Google Maps. But those systems have heaps of prior knowledge about the Underground network; the DNC doesn’t.

The system also carried out tasks such as identifying the relationship of one person to another in a family tree.

DeepMind, owned by Google, has previously developed a memory-aided neural network in 2014 called a “Neural Turing Machine”. The DNC outperformed this “across all tasks”, says Alex Graves of DeepMind.

“I think it’s a very beautiful piece of work,” says Ruslan Salakhutdinov at Carnegie Mellon University’s machine learning department.

Salakhutdinov adds, however, that the system is limited by the size of its working memory: if asked to sort through particularly large sets of data, it may not be able to store all the desirable detail as it searches for answers or patterns.

However, in its paper, DeepMind says processing is independent from the memory, so simply expanding memory size could allow the system to scale up to bigger tasks once trained.

A DNC could be useful in a wide range of applications, Salakhutdinov says, including computer vision in robots. A robot wouldn’t just work out how to open a door, for example; it would learn to keep a record of which ones are locked, so it doesn’t have to keep trying them again and again.