On November 1, 2010, Andrew Herbert, managing director of Microsoft Research Cambridge, was promoted to chairman of Microsoft Research EMEA, and Andrew Blake, deputy managing director of the facility, was named Herbert’s successor. Blake came to the job with a sterling reputation, renowned in computer-science circles for his expertise in computer vision. This article is reproduced with kind permission of Microsoft Research, Cambridge.
1. How is your new role going?
I’m enjoying it. I’ve inherited this highly tuned machine, a sophisticated laboratory with an incredible wealth of talent. I’m still getting up to speed on it all – especially the subjects with which I am less familiar.
I’m spending time getting to know people, and my principal management tool is lunch – sometimes with groups, sometimes with individuals. If I bump into somebody on the way into the lab in the morning, I’ll think, “There’s somebody whom I seldom speak to,” and they get invited for lunch. I just want to get to know people, find out more about what they do. It is a bit like being a headmaster – you should know who all the people in your institution are.
2. Have you encountered any surprises?
I get the sense of a pretty happy ship. The challenge for me is that I’ve been a rather focused scientist, and I only got into management of the lab as deputy director a couple of years ago. I’ve been thinking about my own subject, and now, the challenge is to understand the much broader landscape. There are surprises when I find out what people are doing that I didn’t know about. One area that is interesting is the way in which theoretical computer science and biology are collaborating. There are things coming together that you would never have expected.
3. Can you give an example?
Take Terminator, this utility that Byron Cook has worked on for analysing device drivers and other pieces of code. It turns out that, in a pretty direct way, you can apply that to biological systems. Termination there means death. You want to know if this biological system is going to thrive, or are its metabolic processes going to collapse and the thing dies? It’s amazing how closely coupled the software problem and the analysis-of-life problem are turning out to be. That was a surprise to me.
4. Can you outline your professional background?
I’m coming up on my 12th year in Microsoft Research. I’ve been there almost since the beginning. We were in Guildhall Street in the middle of Cambridge for the first couple of years before we moved out to the university’s science campus. Before that, I was a professor in Oxford on the faculty of the engineering department. I built a research group pretty much from scratch, and we were concentrating on the dynamic aspects of computer vision, seeing things that move.
I got my Ph.D. in Edinburgh, and I was on the faculty there for a while. I had been in the defence electronics industry, working on signal-processing systems for the military.
While in that job, I went to the local university and listened to a man named Joel Trussell giving lectures about image processing. He showed pictures of X-rays, I remember vividly, X-rays of a lung. They were rather foggy, but he treated the X-ray as a matrix and did some matrix operations on it, and the picture came out clearer. I thought it was amazing that, with something as dry and abstract as a matrix, you could have so much fun and make such beautiful pictures. My attention was captured by that.
I never intended to go into research. I thought I would go into a company and work on electronics. But this really piqued my interest, so I started looking for places to get a Ph.D., and I found Edinburgh had one of the pioneering groups in artificial intelligence. I was attracted by this, and I worked for a year or two on code that would achieve what we now call simple edge detection. I took this code down to London and loaded it onto a parallel computer that University College London had developed.
Then I dropped my keys onto the table under the camera, and I saw the edge-detected version of my keys moving around in real time on the screen. The excitement was unbelievable! To see real-time edge detection in 1982 was just breathtaking.
I was inspired by the work of Berthold Horn at MIT, who looked at the problem of analyzing movement in images and the problem of analysing the shape of a shaded surface – how can you tell from the fact that a ball is a shadow on one side and light on the other side that it is, in fact, a ball?
He attacked these problems by using the tools of variational calculus, which is all about optimisation. It was a nice separation of the statement of the problem and the algorithm that would achieve that specification. It was beautiful, the first really interesting approach to machine vision I had seen.
I decided to try to make that work with edge detection, and that was what my thesis was about and, afterward, my work with Andrew Zisserman. That was the first work I did that I really felt pleased about and proud of. I think it’s turned out to be quite influential.
5. How did you end up joining Microsoft Research?
I was in my early 40s, and I had a feeling it would be good to do something new. A lot of my academic colleagues were doing startups, which wasn’t so common then. It was quite brave and bold, and I was full of admiration for these people, but I didn’t quite feel I wanted the distraction of running a company. I wanted to spend most of my time thinking about the science.
When I started doing the thesis work, it sometimes took 24 hours to process a single image. You didn’t think about that being anything like near-term or that maybe one day we’d be able to do this for real. It was just an experiment on how you could do useful computations with images.
By the mid- to late ’90s, things were looking different. We’d already built at Oxford computer systems that would track a moving object in real time. We have just worked on Kinect, and people are getting a lot of fun and enjoyment out of that—the researchers, as well as the users—but in ’94, with my student, Michael Isard, now a researcher at Microsoft Research Silicon Valley, we built a system that could track the movement of your hand in real time, the opening and closing of a hand, and use that to control a three-dimensional object on the screen.
Into that atmosphere was dropped a single, beautifully crafted, two-line email from Roger Needham (founder of Microsoft Research Cambridge), a little missile launched from Cambridge over to Oxford and landing on my desk: “Have you ever thought of coming to work with us?”
When I got the message, it immediately resonated. I was about to go to Redmond to Microsoft to give a talk, but by the time I arrived, the research talk had become an interview. I gave a talk and had a series of interviews with Rick Rashid and Michael Cohen and a lot of people who are leaders in the company now. It didn’t take me long to decide.
We had just built a house in Oxford, and on the one-year anniversary of moving in, we were all out for dinner. Somehow, I mentioned that Microsoft was interested in hiring me. I talked with my wife about continuing to live in Oxford and going to Cambridge a few days a week. She was so horrified at the potential domestic disruption that she said: “No, let’s just move.”
6: What do you see as your priorities—immediate and over the longer term?
The two, in a sense, go together. I’ve inherited an impressive organisation that Roger Needham and then Andrew Herbert have built. Roger built the core of it with the parts that you’d expect to find in any computer-science department in a university. We had a languages group and a systems group and then a machine-learning group.
Andrew added two groups that are a little more adventurous. There was the Computer-Mediated Living group. That’s our name for human-computer interaction, but the name is different for a good reason: The way we approach it is rather different, with an emphasis on design rather than simply on analysis and with an interesting component of anthropological expertise. It’s quite an unusual mix.
The other was the last group to be added, the Computational Science lab, where we’re doing research on computational ecology and computational biology. Of course, when you look at the core of Microsoft’s business, it’s a much bigger leap from traditional products and services to this. But interestingly, linkages are emerging – and some ambitious technologies, such as computing with DNA. We have people getting strands of DNA synthesised. The strands behave like electronic gates, “and” gates and “or” gates.
It’s just incredible. We’re a long way from knowing exactly what kind of machine we would want to build to exploit that, but these computational phenomena are so fascinating they have to be investigated, and it’s great that we’re sufficiently broad to encompass that.
I have an aspiration that, in five years’ time, I’d like to have five great new stories to tell, five substantial contributions that mean something to the company and to society. But it’s a delicate balance. The culture of the lab is that it’s driven by research and driven by people’s scientific interests, yet also the lab is regarded as a machine, and we want the output of this machine to be highly influential in the company. I’m already speculating about what my five big stories might be.
7. Are you talking about contributions similar to those Microsoft Research Cambridge made to Kinect?
Kinect is a great example. It might be more of a research phenomenon, where the chemistry is just right between what the research group is doing and what the product group wants, resulting in a marketplace impact that is gratifying. I think of F#, Don Syme’s work, as being another big project. Developer tools are an area where Microsoft absolutely excels, and F# is a jewel in that crown. That’s something that our lab is very proud of.
Infer.NET may have similar potential. We’re only just beginning to understand its power. It builds on deep research that has been going on for at least a decade by Chris Bishop, John Winn, Tom Minka and John Guiver, about how to model probabilistic processes. Infer.NET is an embodiment of that as a set of tools, which makes them much more influential. People can actually pick them up and use them.
With Infer.NET, the challenge is to bridge the gap between something that requires specific machine-learning expertise and the skills that developers bring to bear. We have to encapsulate the ideas in the right way and build the right tools.
8. As managing director, will you still have time to pursue your own research?
Yes – not as much time, of course, but it’s going to be helpful to continue to be immersed in research so that I continue to understand what it’s like to do leading-edge research and so that people in the lab relate to me as a researcher. It’s tremendous fun doing research. It’s what I’ve done for the last 30 years. I’d be loath to let it go.
Some of the people I hired right after their Ph.D.s have grown into strong and independent researchers with substantial international reputations. I am more at arm’s length now from their day-to-day work, but I’m grateful that they still let me join in.
Here are two research interests I’m pursuing now. One is with natural user interfaces, which I think about as an evolution: First, we had green screens and cursors, then we had the mouse and windows, then we had touch and multi-touch, and now we have no-touch: action at a distance.
What are the no-touch interactions going to be like? Not just with computers, but with machines that don’t look like computers at all but have embedded computation, things like automobiles, DVD players, maybe microwave ovens. There’s a potential to interact with all of these things at a distance. I’d love to see us provide a middle layer to enable creative people to design their user interfaces with our technology. Another is an interaction with chemical engineers at the University of Cambridge. We’re looking at magnetic-resonance imaging (MRI) in a completely new way.
At the moment, magnetic-resonance imagers like the ones in hospitals use a big, super-cooled magnet. You can’t move them, and they cost a fortune. We’re trying to use Bayesian inference to determine directly the properties you’re interested in for a particular system, bypassing the formation of an image . This could be used with the human body, or for something in chemical engineering. The people I’m collaborating with are interested in MRI as a tool in analysing chemical processes. It turns out you can do that incredibly efficiently.
You could make machines where the magnets are not so powerful, you no longer need super-cooling, and you can use a big permanent magnet. Then we’d have a machine that is portable. This research is in its early stages, but we’ve proved that you can do this and that the theory works. In the most optimistic scenario, we could define completely new kinds of imaging machines that would be much easier to use.
9. Can you tell us something about yourself that people would be surprised to hear?
Well, there is nothing so surprising. I play the violin. I love music, and I’m a rank-and-file violinist. I love to play choral music, especially Bach. I did try singing recently, the Fauré Requiem, but I decided that I probably had better go back to playing the violin.
I also like hiking. The Lake District in England is my favourite spot, and I go back there quite a lot. It’s just a beautiful place. The thing that’s special about it is it’s not just countryside, it’s also the way the built environment is nestled into the countryside—all of those beautiful old villages with their granite and slate, dry stone walls without any mortar. I love that combination.
10. What about Microsoft Research Cambridge makes you the proudest?
An amazing bunch of people. What more can I say? Just incredible people. I’m constantly amazed at the quality of people you meet and their passion for research – and, to be honest, what a nice bunch of people they are. It’s a lovely atmosphere, a place where you can follow your passion for research.
Many academics would dream of such an environment. It’s a very uncluttered space. Microsoft is great at taking care of the earthly and temporal matters so that you can focus on the research. I see a whole bunch of people in very diverse disciplines who are absolutely passionate about doing that.