Teaching computers to think and play games
John Murphy talks to Christopher Bishop, assistant director of the Microsoft Research Laboratory, Cambridge
Whatever people might think about the business methods of John D Rockefeller, nobody today questions the contribution of his legacy to medical science. Similarly, whatever the academic community thinks of the works of Microsoft, there is no question that the company has sponsored some good research.
The laboratory it founded in Cambridge, England, in 1997, is the base of one of the leading researchers in machine learning, who has applied his knowledge of theoretical physics and mathematics to make significant advances in the field. Professor Chris Bishop has also formed a group that not only publishes significant research work, but has also devised an artificial intelligence engine which will soon be appearing in Microsoft games and which will be used by millions of people. Although the group was not set up to do product development, it just seemed like the kind of application that would test the practical potential of a research idea. Bishop's group has also made public-domain tools that make neural networks easier to use.
Professor Geoffrey Hinton of Toronto University, a collaborator with Bishop, said it was a shame that universities could not afford to pay people as much as Microsoft. Maybe as a result, Bishop had built up an excellent research group doing important work that was being published to the academic community.
He said that Bishop really made his name with a textbook on Neural Networks and Pattern Recognition, published in 1995, which had for the first time explained the links between neural networks and their basis in the mathematics of statistics. His recent publications while at Microsoft Research, Cambridge had continued to have a strong influence. Professor Hinton said: 'Chris is extremely strong on the mathematics, but he does not have his head in the clouds. He likes to see things applied to real world problems, and he has worked on some of the leading application of neural networks. He is certainly very determined and likes to get the best out of everything.'
Bishop was born in Norwich, in the east of England, and from his childhood knew he wanted to grow up to be a scientist. In fact, while at school, he built his own CO2 laser by scrounging components from various companies and university laboratories. His nickname at school was 'professor'. He did well at school and won a place at Oxford to study physics.
During the holidays he worked, but did not choose the normal options of shops or fast food restaurants. His first summer job was at the Propellants, Explosives and Rocket Motor Research Establishment (part of the UK Ministry of Defence and formerly known as the Royal Gunpowder Factory), which he described as terrific fun. He worked on the use of holograms to do non-destructive testing, but there were plenty of fireworks. He said: 'We used to schedule the firing of solid rocket motors at lunchtime, so that everyone could watch them as they ate lunch.'
His second summer job was at the Culham Laboratory, a government research station that he had heard about because of its Joint European Torus (JET) nuclear fusion programme which had just been announced. He worked as an experimentalist, even though he was still an undergraduate.
After graduating, he decided to go into theoretical physics, and applied to Edinburgh University because he wanted to study under David Wallace, who had worked with Nobel Laureate Ken Wilson in the US. Wallace was setting up a new group at Edinburgh and Bishop was his first PhD student, although he was also taught by Peter Higgs (of the eponymous particle). His PhD was on the use of semi-classical methods in quantum field theory, but at the end of it he decided he really wanted to do something practical. So he took a job back at the Culham Laboratory as a theoretical plasma physicist. His work was on the H Mode of plasma in high confinement, which had only just been discovered through experiments.
He said: 'I happened to be working on the magnetic configurations in which this H Mode was seen to occur, and so I guess I was lucky in that I wrote the right paper at the right time, in which I gave an analysis of what was going on. I had only been there a couple of years and was unheard of in the field. Then, suddenly, I was invited everywhere to give talks on this paper. Some of the predictions of the theory later turned out to be confirmed by experiment. It was very nice to go completely around the scientific process from experiment to theory and prediction to experiment.'
Bishop had always been fascinated by the human brain and the idea of artificial intelligence. He had some contact with the AI group in Edinburgh, but was not happy with the rule-based, expert-systems approach which it was working on at the time. He was much more interested in the idea of building an adaptive machine that could train itself to solve problems. By the mid 80s, the AI world had moved on and he became interested in working in the field again. He had by this time moved to AEA Technology, which was looking to diversify its research and so he suggested using machine learning as an approach to machine diagnostics in the world of fusion. In the end, he became so fascinated by machine learning that he decided to specialise in it and was head-hunted for a position at Aston University, where a new Neural Computing Research Group was being established and, fulfilling his school friends' prediction, he became a real professor.
At Aston, he developed many new interests and started working with industrial partners on projects, but he also worked on the theoretical basis of neural computing and he developed new general algorithms. In 1997 he was granted a sabbatical from Aston to pursue a research project at the Isaac Newton Institute for Mathematical Sciences in Cambridge. He spent the next six months working there and met hundreds of people working in the field of machine learning through its programme of seminars and workshops. One of the people he met at Cambridge was Roger Needham, who had run the widely respected Cambridge Computer Laboratory for many years and had just set up Microsoft Research, Cambridge.
The establishment of Microsoft Research in Cambridge had caused a degree of controversy in the academic community at the time. Needham had to establish its academic credentials as quickly as possible, and offered Bishop one of the top jobs there. Bishop joined, but was allowed to continue with his project at the Isaac Newton Institute. Needham died about 18 months ago and Bishop has edited a collection of his monographs as a tribute.
At the same time, Bishop had been talking to people in the computer science department at Edinburgh and, until the Microsoft job came up, had expected to end up moving to a chair in artificial intelligence. He said: 'I knew that the offer to work at the Microsoft lab was an opportunity I could not possibly turn down, but I was sad not to be going to Edinburgh. I had a difficult phone call to make to the principal, Stuart Sutherland, to say that I would not be moving to Scotland. He had a marvellous suggestion: why don't you do both jobs? In practice I go to Edinburgh about once a month for a few days. In research it matters who your colleagues are, so having the joint appointment means that I can spend time working with colleagues in Edinburgh, which is very useful.'
Bishop has no regrets about joining Microsoft. Unlike in the academic world, where so much time is taken up applying for research grants, the laboratory has a budget and he and colleagues are able to decide how it is spent. In particular, there are not the same restrictions on travel that tight academic budgets impose. Most importantly, he has been able to continue publishing his research in the same way as if he was in a university position. He said: 'Since I have been here, I don't think I've had to turn down a request to travel to a conference.'
His recent research, in collaboration with John Winn, has been around variational inference, which uses approximate Bayesian techniques for inference. They have developed a way of describing problems as probabilistic graphical models - which involves drawing a picture of the problem on screen with a mouse and the tool generates the software necessary to implement it. The research has just been accepted for publication, and the software tool has been published on the web for others in the field to try it out and give feedback.
His group has also surprised everyone by developing an AI engine to run a forthcoming XBox motor-racing game called Forza. The traditional approach is to use a rule-based system with a probability element - but the user does not feel as if they are playing against a human player. Bishop's group has trained a machine-learning programme with live data from a human racing driver. The lab was never intended to do product development, and Bishop said it was a struggle to get the product development team at Microsoft to accept something from an academic team. He said: 'The work did not come about because anyone asked us to do it; my group were keen on playing computer games and decided they could use machine-learning to make computer games. So instead of publishing a paper, they will have their research in a game used by millions of people. But there isn't a formal process for bringing research and product groups together. They had to go to Redmond and do a hard-sell job, because not only were they doing something new with research that had to fit into a product development cycle, with the associated risks, but also we were proposing something very different. We had to work hard to persuade the product groups that the risk was worth the possibility of using new technology.'
Bishop believes the computers of the future will be much more adaptive. He said: 'I think that, at the moment, computers are incredibly dumb. The computer is there as a tool, but if you compare a computer as an assistant to a human as an assistant then you find that on the first day they can't help much but, after a few days or a few weeks, the assistant would be able to help more. Computers don't change and adapt and learn to work with you. One of the long-term goals I have is to learn how to make computers that adapt themselves to your personal needs.'
When he is not working, and sometimes when he is, he is also a keen pilot. For fun, he flies 1930s biplanes. While most people would be happy with a Private Pilot's Licence, Bishop decided that if he was going to be flying his family around and using a plane for business, he ought to get a full Air Transport Pilot's Licence. Friends believe this is an indicator of how determined and thorough he is, as it takes many years and many hundreds of flying hours to gain this professional qualification.