Self-programming computers (little fake brains) fuel new start-up.

You know it’s real when there’s money involved. Well, real-ish. New Scientist has more on the Google acquisition of DeepMind Technologies and their Neural Turing Machine:

DeepMind Technologies, a London-based artificial-intelligence firm acquired by Google this year, has revealed that it is designing computers that combine the way ordinary computers work with the way the human brain works. They call this hybrid device a Neural Turing Machine. The hope is it won’t need programmers, and will instead program itself.

“These neural networks that are so good at recognising patterns – a traditional domain for humans – are not so good at doing the stuff your calculator has done for a long time,” says Jürgen Schmidhuber of the Dalle Molle Institute for Artificial Intelligence Research in Manno, Switzerland.

Bridging that gap could give you a computer that does both, and can therefore invent programs for situations it has not seen before. The ultimate goal is a machine with the number-crunching power of a conventional computer that can also learn and adapt like a human.

DeepMind’s solution is to add a large external memory that can be accessed in many different ways, which mathematician Alan Turing realised was a key part of ordinary computing architecture, hence the name Neural Turing Machine (NTM). This gives the neural network something like a human’s working memory – the ability to quickly store and manipulate a piece of data.

To test the idea, they asked their NTM to learn how to copy blocks of binary data it received as input, and compared its performance with a more basic neural network. The NTM learned much faster, and could reproduce longer blocks with fewer errors.