An Interview with Dr Ruediger Urbanke, Professor at the EPFL Institute, Switzerland
Dr Ruediger Urbanke is the recipient of the IEEE Richard W Hamming Medal 2014. His current research is sponsored by the Swiss National Science Foundation, the NCCR/MICS (National Center of Competence in Research / Mobile Information & Communication Systems) as well as CTI, Switzerland.
Q 1. Please introduce yourself for our students.
My name is Ruediger Urbanke. I am a professor in the Department of Computer Science and Communications at EPFL, Switzerland. I have spent more than 25 years studying error correcting codes. I was lucky to do this at a time when many exciting new developments happened in the field. If I am not thinking of codes, I like to work on my bicycles. My current project is to build a foldable tandem from scratch.
Q 2. What kind of environment in your childhood motivated you towards electrical engineering? Did you involve yourself in any hobby projects like taking a peek inside a radio or a television and so on?
Starting from early on, I took many things apart, not necessarily electrical. Few of these I even put back in working order, however. That I studied EE happened somewhat randomly, probably mostly due to the suggestion of my parents. I had a very vague idea what EE was really about. In fact, when I initially contemplated studying EE, my biggest concern was that I might not be able to determine the impedance of a given resistor due to my partial color blindness. Needless to say that this problem never came up. So in short — I was quite naïve and randomness played an important part, but it worked out very well.
Q 3. Electrical, Electronics and Communication is an engineering domain with a vast number of subdomains in it like VLSI, microwave engineering and so on. What incidents, events or circumstances made you choose information coding, transmission and decoding as your domain for exploration after graduating with an electrical degree from the Technical University of Vienna in 1988?
At that time in Austria, the study of EE was split into two parts, a 3 year general part, followed by a 2-year specialization. I liked math from the very beginning, and when it came time to decide which branch to follow, it was crystal clear that I wanted to do communication. It was just about the time where communications made a rapid phase transition from analogue to digital. And from anything I had seen about digital communications, it had a basis principled in math, it was tremendously useful, and there was a sheer endless list of interesting problems to work on.
Q 4. How was your experience in the Bell Laboratories during the period of 1995–1999? Did it motivate you to pursue a career in iterative decoding techniques?
During my PhD, I worked only peripheral on coding, although I had always a keen interest in this topic. In fact, I never took a formal course on this topic. So when I joined Bell Labs, I was initially supposed to work on wireless and information theory. But it was the period shortly after Berrou et al had discovered turbo codes and so there was this great practical breakthrough that was looking for a theoretical explanation. It was comparable to what you see these days in neural nets. Many people hence started to think about this problem. During several years, Tom (Richardson) and I would talk daily for hours to see if we could make progress on this topic. These were really exciting times.
In addition, Bell Labs had so many outstanding people. Whatever you wanted to know, there was a good chance that in one of your neighbouring offices you could find somebody to explain it to you. Many of my best friends and closest collaborators are still from this period.
Q 5. Is communication engineering all about mathematics? Please let us know your opinion on this.
Math is an integral part. But what I love about this topic is that there are few engineering disciplines where theory and practice are so closely linked and where both intuition and fundamental math go so well hand in hand. Over the years, sophisticated algebra, number theory, probability theory, and ideas from computer science and statistical physics have all influenced how we view and design codes. And many of these ideas made it in a relatively short time period from theoretical papers into actual products.
Q 6. How do you think focusing on better error-correcting codes can help in the development of algorithms for artificial intelligence and self-driving cars?
Coding has been declared dead many times, only to come back with a vengeance every time. There is always a new angle to the problem, new applications, new boundary restrictions. Although I personally have not worked on these problems, I have seen many interesting instances of how ideas from coding can be helpful besides the basic transmission problem. It has been shown that if we do not store random patterns, but patterns with redundancy in a neural network then the capacity of the network goes up dramatically. Coding can be used for distributed computing applications where a computation is split into many small parts that are each performed by different machines. If you add redundancy to the computation you do not have to wait for the last of these computers to return its result, significantly reducing waiting times. In addition, many of the ideas that have been developed in the framework of coding are well suited for problems in machine learning. Just think of message-passing algorithms and how they can be used for optimization problems. There is no doubt in my mind that we will see much more of this in the near future.
Q 7. The LDPC codes discovered in 1963 were forgotten with time and regained importance with the discovery of Gallager’s works in 1996 when Turbo codes were popular in satellite communication. What made you believe that LDPC codes can make a breakthrough in data transmission and pursue a research career in it at a time when Turbo codes were very popular?
My main motivation was not the design of particular codes but rather I was interested in explaining why things work. So it mattered little to me if LDPC codes were better or worse than Turbo codes. It just turned out that LDPC codes were slightly simpler to study. And this is why much of my work with Tom Richardson was focused on LDPC codes. But in hindsight the same ideas that explain why LDPC codes work also apply to Turbo codes.
Q 8. Given your experience on working in the mathematical department of communication systems in Bell Laboratories, what made you to believe that physics and theoretical computer science can usher in a new era in the development of error-correcting codes? Was there any specific incident or event in your life associated with that stance?
It just happened that several people studying iterative coding in the late nineties come from these areas. Naturally, I and others in our community would talk to them. This took many years and many workshops. The differences in language various communities use can be quite a hurdle. Eventually, from these collaborations and interactions, some very fruitful ideas emerged. But at the beginning, it was far from obvious if this would lead anywhere. The lesson for me from all of this is that in such situations you have to take a leap of faith and hope for the best. Worst case, you learn a lot. And if you are lucky, something nice comes out of such collaborations. Don’t try to plan everything. Often the most interesting ideas are the results of random interactions.
Q 9. How have you planned to contribute to your current interest area of framing computer and communication problems as inference problems on sparse graphical models? How is your previous work helping you in exploring this field?
I typically do not have long-term plans. I look for problems that are interesting, fundamental, and have a potential practical impact. Over the years I have come across many ideas and areas and sometimes these ideas, later on, find applications beyond their original intent. E.g., many years ago we were interested in the question of how well an optimal decoder (not the message-passing one) would do if used together with an LDPC code. This seemed a very natural and fundamental question although it did not have any imminent practical application, since such an optimal decoder is, in general, not feasible to implement. But it turned out that a few years later this was exactly the missing link that explained why the so-called spatially coupled codes work so well under iterative decoding. There is no way we could have planned this.
Q 10. Your advice for students taking an undergraduate degree in electrical, electronics and communication engineering? What are the prerequisites a student needs to pursue a career in the domain of coding and decoding message signals in noisy channels?
The main pre-requisite seems to be the ability to get excited about this topic and a good background in math. Be open to new ideas. Many of the innovations in our field have come from outside the traditional community and by the same token, ideas from coding are bound to have many more applications in other areas. Look for these opportunities!
Further Information:
- https://people.epfl.ch/rudiger.urbanke/
- https://ethw.org/R%C3%BCdiger_Urbanke
- https://en.wikipedia.org/wiki/Low-density_parity-check_code
This interview was taken in collaboration with Saran of the Probe Content Team. We would like to thank Dr Ruediger Urbanke for taking out time to share his words of wisdom and experiences with the student community.
Clap and share if you liked this one and do follow Probe, NIT Trichy for more insightful updates from the world of electronics!
