Can Google’s AlphaGo really feel it in its algorithms? | John Naughton

Last week, researchers at the artificial intelligence company DeepMind, which is now owned by Google, announced an extraordinary breakthrough: in October last, a DeepMind computing system called AlphaGo had defeated the reigning European champion player of the ancient Chinese game go by five games to nil. The victory was announced last week in a paper published in the scientific journal Nature.

So what? Computers have been getting better and better at board games for yonks. Way back in the dark ages of 1997, for example, IBM’s Deep Blue machine beat the then world chess champion, Garry Kasparov, at chess. So surely go, which is played not with six different pieces but black and white tokens – would be a pushover? Not so: the number of possible positions in go outnumber the number of atoms in the universe and far exceed the number of possibilities in chess.

Illustration by Matt Murphy.

Game-playing programs such as Deep Blue work by constructing “search trees” over all possible positions. If you tried to do this with go you would be looking at computing times in the millions, if not billions, of years. The DeepMind approach involved combining tree search (enabled by Google’s massive cloud computing power) with deep neural networks, programs in which connections between layers of simulated neurons are strengthened through examples and experience (as in the human brain).

AlphaGo first studied 30m positions from expert games of go, absorbing information on the state of play from board data, and then played against itself across 50 computers, improving with each iteration, using a technique known as reinforcement learning. It’s a development of an experiment DeepMind used a while back, in which one of its neural networks taught itself to play retro computer games to a proficient level merely by watching replays of real games on a screen. No programming involved.

The really significant thing about AlphaGo is that it (and its creators) cannot explain its moves. And yet it plays a very difficult game expertly. So it’s displaying a capability eerily similar to what we call intuition – “knowledge obtained without conscious reasoning”. Up to now, we have regarded that as an exclusively human prerogative. It’s what Newton was on about when he wrote “Hypotheses non fingo” in the second edition of his Principia: “I don’t make hypotheses,” he’s saying, “I just know.

But if AlphaGo really is a demonstration that machines could be intuitive, then we have definitely crossed a Rubicon of some kind. For intuition is a slippery idea and until now we have thought about it exclusively in human terms. Because Newton was a genius, we’re prepared to take him at his word, just as we are inclined to trust the intuition of a mother who thinks there’s something wrong with her child or the suspicion one has that a particular individual is not telling the truth.

The trouble is that intuitions can be – and often are – wrong, which is why we demand evidence and reasoned argument to support them and why we become suspicious when these are not forthcoming. For the time being, concern about machine intuition is moot: it doesn’t really matter that AlphaGo cannot explain the basis for its intuitive moves in go. But that won’t last if Dr Demis Hassabis, the co-founder of DeepMind, has his way. “While games are the perfect platform for developing and testing AI algorithms quickly and efficiently,” he wrote in a blog post last week, “ultimately we want to apply these techniques to important real-world problems. Because the methods we have used are general purpose, our hope is they could be extended to help us address some of society’s toughest and most pressing problems, from climate modelling to complex disease analysis”.

On the one hand, Hassabis seems to appreciate the responsibilities that those who create intelligent technologies will have to shoulder. He’s been calling for “a responsible debate about the role of ethics in the development of artificial intelligence”. On the other hand, when pressed about this at an event in Cambridge last September, he tried to fob off the questioner with the news that Google was setting up an ethics committee to look at the work his company is doing, a response that evoked a certain amount of hilarity in some sectors of the audience.

Like most people in his field, Hassabis thinks that AI powerful enough to pose serious ethical and existential threats is a long way off. The remarkable progress his company is making, however, suggests otherwise.

comments powered by Disqus