Google seems to be serious about getting a slice of the AI business. It has now bought the UK start-up DeepMind Technologies for a very large sum of money, even though the company hasn't actually done anything very much yet.

DeepMind was founded in London in 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman.

Demis Hassabis is the one of the three with a larger than life persona. He was a child chess prodigy and has been involved in games of all types ever since. Working in the games industry, he created some well known games such as Theme Park. He then  studied computer science at Cambridge and went to on to get a PhD in Neuroscience from University College London in 2009. His best known research isn't so much in the area of AI as experimental neurology - a study of the way that amnesia due to damage to the hippocampus effects the ability to imagine the future.

The company has very little public presence but its home page says:

"DeepMind is a cutting edge artificial intelligence company. We combine the best techniques from machine learning and systems neuroscience to build powerful general-purpose learning algorithms."

and

"Our first commercial applications are in simulations, e-commerce and games."

That's the only clue the company provides as to what it does and what approach it takes.  Another clue is the fact that it seems to have hired recent graduates from the best known deep learning research groups - those led by Hinton, Lecun, Bengio and Schmidhuber. Google has already hired Hinton, part time at least, and has been focusing on research using deep neural networks.

Perhaps the best clue as to what might have interested Google in the company is a recent paper published by a team working at DeepMind that demonstrates using a deep neural network to play a range of Atari 2600 games, including Breakout.

What is novel about this work is that it doesn't use supervised or unsupervised learning to train the network but reinforcement learning. The network is trained using a modified form of Q learning which is about the only valid approach to machine reinforcement learning. In this case the machine isn't told how well it is doing at each move or just left to find patterns in the data it learns according to the reward it gets during the game and particularly at the end when it gets a win/loss reward. 

Could it be that Google hopes to add the reinforcement learning method to its deep neural networks? If so the cost of the company, estimated at $400-500 million, seems small  - but only compared to the crazy $3.2 billion it paid for the AI thermostat maker, Nest. It has also been rumoured that in buying the company, Google beat Facebook, which recently recruited well-known pioneer in the field of machine learning and computer vision, Yann LeCun.

Well will have to see if deep reinforcement learning makes an appearance in any future Google AI projects. But it seems much more likely that this was just a purchase of personnel and their likely role is to see how AI can improve search.