Advanced computer can learn best route by trial and error

System analyses data by blending neural network and standard computer memory

The system “can successfully answer synthetic questions designed to emulate reasoning and inference problems in natural language”, say scientists. File photograph: Getty Images

The system “can successfully answer synthetic questions designed to emulate reasoning and inference problems in natural language”, say scientists. File photograph: Getty Images

 

Computer scientists have come up with a new kind of learning computer that can navigate a map of the London underground and choose your best route.

It blends two types of machine, a neural network and standard computer memory, allowing it to analyse data and learn through example or by trial and error.

The work is under way at the Google DeepMind research centre in London, so what better an example to begin with than the underground network just outside.

When a human studies a map of the London underground we intuitively look for our starting and finishing point and the train lines connecting these points.

This requires us to use pattern recognition to understand the connections and also memory so that we can hold the required route as it is being assembled.

Artificial neural networks are good at identifying patterns but lack the memory systems needed to process the data.

Alex Graves, Greg Wayne and Demis Hassabis and colleagues got around this by linking a neural network which can learn as it goes with an external memory system similar to standard random access memory in any computer.

Prior knowledge

The result – a “differentiable neutral computer” – can understand graph structures like the underground map, other transport networks or even family trees.

It can plan the best route from point-to-point across London without prior knowledge of the system. And it can learn to solve moving block puzzles just by analysing an explanation of how the game works, say the scientists in the journal Nature published on Wednesday evening.

The neural network is supported in this by the attached memory, which processes the complicated data like a computer.

“Like a conventional computer, it can use its memory to represent and manipulate complex data structures, but, like a neural network, it can learn to do so from data,” the authors write.

After supervised learning the system “can successfully answer synthetic questions designed to emulate reasoning and inference problems in natural language”, they say.