The trouble with deep neural networks

Fan Hui getting defeated by AlphaGo

In October 2015, and in relative secrecy, a major milestone in Artificial Intelligence was reached. AlphaGo, an algorithm developed by Google to play the game Go, beat Fan Hui, the Go champion of Europe, by five games to zero. The results, as well as an insight to the algorithm, were published in the scientific journal Nature early this year.

There are several reasons why Google’s achievement is significant. First, it’s the huge degree of complexity of the problem they tackled. The game of Go is played by two players on a 19×19 grid board, the goal being to gain as much territory as possible by moving and capturing black and white stones. The average 150-move game contains more possible configurations that there are atoms in the universe (10 followed by 170 zeros, to be exact). This so called “search-space” is simply intractable for conventional computing approaches.

So the second significant reason for paying attention to what Google did – or rather DeepMind which is Google’s AI company – was their approach to solving the search space problem of Go. They expanded on techniques that they had already developed, and which helped them win in the 49 arcades games early last year. And they put those techniques on steroids using a combination of deep neural networks, supervised learning, and reinforcement learning. They started out by training their networks to mimic the moves of human players using 30 million positions from games played by human experts. Machine learning is all about applying the mathematics of probability theory to track how closely a computer can “guess” the correct output when given a certain input. Training – or “supervised learning” as it is technically called – entails the computer analyzing massive data sets and building a probability function that describes relationships between them. It then applies this function to any new data set. Once supervised learning was complete the scientists at DeepMind let the algorithm play against itself across 50 computers. This “reinforced” its ability for pattern recognition, improving its game with each iteration. Finally, they added a search approach with the ability to pick moves and interpret Go board, thus furnishing AlphaGo with the ability to rank the strategies that were most likely to succeed at any given move. The result was a program that is scales of magnitude better as playing Go than any other computer program before it, and better than an expert human too. As a follow up test of AlphaGo’s ability, a match between it and South Korean Lee Sedol, the world’s most titled player, is scheduled for March 2016.

The achievement of AlhpaGo was heralded in Nature, and other media outlets, as the first sign of “machine intuition”. If this is true, then it makes DeepMind’s technological breakthrough all the more significant. By intuition we usually mean that uniquely human “eureka moments”, when a true statement reveals itself to us not through logical reasoning but out of nowhere. Intuition is fundamental to the sciences and arts, and has been studied by historians, psychologists and philosophers. It seems that all big human discoveries have been made by intuition. The experience of intuition is that, because it is not the result of a sequence of logical steps, one can never explain how that eureka idea came to them. It is only by hindsight that scientists will go back and work out the mathematics and the logic behind a great, intuitive idea. So does AlphaGo match this human capability? And if so to what degree?

Let’s compare AlphaGo with Deep Blue, the other artificial intelligence computer that made the news back in 1997, when it beat at chess the world’s chess champion, Gary Kasparov. The big difference was that Deep Blue’s program was handcrafted. In other words, it was possible to go back to the source code and trace why Deep Blue made the moves that it did, or why it followed a certain strategy . Its behavior was rational, in the sense that humans could query the machine and get an explanation of its reasoning. AlphaGo is a completely different technological beast. It uses “deep neural networks” which is a metaphor for computer systems that self-organize through experience and learning, rather than process data through hardcoded facts and descriptions. Because of self-organisation these systems are very efficient in identifying patterns in complex data. Indeed it this capability of deep neural networks that makes them such an exciting technology. When the decision-making task is challenging, the search space intractable, and the optimal solution complex, deep neural networks are the answer. The problem is that <em>the answer</em> that the neural network will deliver cannot be queried, but has to be trusted and believed.

Having to place blind trust on a machine’s judgement is something completely new for human civilization. It is a challenge of enormous ethical, political and moral dimensions. Take for example medical diagnosis. Many years ago I had built an expert system for medical diagnosis that used a handcrafted approach, similar to Deep Blue. A fundamental requirement for the acceptability of that expert system in any real medical environment was that the system ought to give an explanation of its diagnosis. Human doctors need to explain why they think a patients suffers from a certain disease. A diagnostic system based on deep neural networks will probably be a lot better that any expert system or any human doctor could ever be. It will deliver the best and most accurate of diagnoses, but it could not explain the reason why. The experience of communicating with such a machine would be similar to communicating with an oracle: “truth” would appear to come from nowhere.

Perhaps the success of deep neural networks in years to come will be so spectacular that we humans will adjust to simply trusting their judgment, unquestionably. But what an irony would that be! For science and technology, the twin siblings of rational thought and the scientific method, will have given birth to something that resembles blind faith and religion. If deep neural networks are the future, then the future will look strangely like the past.