An interview with Sébastien Destercke
Posted on July 7, 2022 by Sébastien Destercke (edited by Henna Bains)[ go back to blog ]
This blog is the first instalment in a new blog series where we interview members of the SIPTA community.
A few weeks ago, Henna had a conversation with Sébastien Destercke about his career so far, thoughts on imprecise probabilities, and more. Please enjoy the interview below.
To start, could you give the reader a quick insight into who you are by describing yourself in three words or phrases?
This one is difficult; hmm, I’d say jack of all trades, gamer and conversationalist (I like to engage in long conversations about topics of interest).
Could you describe your career so far?
How did you get to where you are now?
I attended an Engineering school in Belgium, in a town called Mons (or Bergen in Dutch), and I did my last year in Supaero (an Engineering school in Toulouse, France). It is in my last year that I got in touch with research; I did not have an exact idea of what I wanted to do before that. Hearing researchers talk about their work stirred my interest, as I thought what they did was fun, so I started searching for research internships.
I found one in Montpellier, which was about inducing fuzzy expert systems from data. It was connected to a thesis proposal that I didn’t get, but that allowed me to get in touch with Didier Dubois. He made quite an impression on me. The internship also allowed me to confirm that I liked research very much. Later on, I asked Didier if he had other possible topics for a thesis, and that’s how I started my PhD studies at an institute working on nuclear safety. The topic concerned uncertainty reasoning with expert opinion and performing risk analysis for nuclear safety issues when there is little information, for example, to assess the potential of the nuclear reactor overheating. Under the advice of Didier, I naturally focused on imprecise probabilities to handle this little information. Computations were quite an issue then, as a single run of the computer code simulating the nuclear reactor could take a week.
After that, I found a position as a Research Engineer at the Institute for Agronomical Research in France, that was connected to the people with whom I did my internship. I continued doing uncertainty reasoning for two and a half years, but this time to aid decisions applied to agronomical issues. For example, one application looked at how you can design packaging for vegetables and fruit to remain fresh for as long as possible on the shelf. Working there made me realise that I was more interested in focusing on conceptual and theoretical research to apply them from time to time, rather than working on applied projects and from time to time looking at theoretical aspects.
This is why I applied to CNRS, the French National Research Centre, where you can get a permanent research position with many research freedoms. I applied to Heudiasyc, where I am working now. It’s been around 11 years that I’ve held this position, and it’s very nice here. I’m still working on concepts and theory, and I like applying things from time to time. Currently, I’m working on a variety of topics, with a focus on machine learning and inference in combinatorial spaces, and applications ranging from autonomous vehicles to plastic sorting.
It was quite a journey with many turns!
Can you take us back to the beginning - When and how did you start working in imprecise probabilities?
My very first encounter with imprecise probabilities was during the internship I mentioned earlier. In 2004 I had the opportunity to participate at a conference called LFA (French conference on uncertainty theories), in Nantes. I had to present my first paper (I was not yet a PhD student) with my supervisors, and I was very stressed! One of the invited speakers was Gert de Cooman. This was the first time I heard about imprecise probabilities, and I must admit that I understood at most 5 to 10% of the talk at the time. Gert was a wonderful speaker, but I knew nothing about imprecise probabilities back then.
After that, I’d say my encounter with Didier. How he engaged me with imprecise probabilities, I would say, is by telling me that there is a researcher called Arnold Neumaier who proposed an uncertainty model called Cloud. Didier felt clouds were very much related to possibility distributions and things like that; in particular, Didier felt this was a special case of belief functions and asked me to prove that. So I got on with the task, working sometime during train travels on the proof and counterexamples. It was actually very fun. I discovered that only some types of those clouds were actually belief functions (and some were not). You need an extra property for them to induce belief functions. I was proud of myself. Didier was also happy that we sorted these things out.
This led to many other nice things, such as collaborating with Enrique Miranda and Matthias Troffaes around the notion of p-boxes and generalised p-boxes. This was the start of my interest in imprecise probabilities and many, many nice collaborations.
So, this is how it started, and it hasn’t stopped since then.
We have touched on this already, but could you tell us a little more about the aspects of imprecise probabilities that you have focused on?
As I said in my initial three words, I’m a bit of a jack of all trades, so I like to look at quite a lot of things. If you look at imprecise probabilities in the broad sense, the things I’ve mostly focused on are practical representations, that is special cases of lower probabilities and characterising their pros, cons, and properties. This is still something I do from time to time right now. Our recent work with Enrique Miranda and Ignacio Montes on distortion models is a good representative of this strand of research that has been going on for a long time. If I have the opportunity, sometimes I use those practical representations in some practical applications.
Another one that I’ve focused on since my thesis and is still an interest of mine right now is the topic of information fusion, especially the idea that you should allow for conflict (or incoherence) in such a case. I think measuring and handling conflict is interesting in a bunch of situations, and this is something that the belief function community has been focusing on for quite a while. Our work with Thomas Burger and Frederic Pichon typically follows this trend.
More recently, I’m trying to focus on making contributions in fields where I think imprecision can be helpful. For example, active learning in machine learning. Basically, trying to include IP or imprecision in other fields … so again, my jack of all trades coming in.
What would you say you enjoy most about what you do?
It is difficult to identify only one item, as I enjoy many aspects of the research activity. One thing I enjoy very much is learning things and the process of discovery. The procedure of finding an interesting question and the process of finding an adequate solution gives you a very satisfying feeling. I also enjoy connecting things together, building bridges.
Meeting and discussing with others is also something I like very much, as research is an area where you can meet such fantastic and fascinating people and ideas (more than in other areas I think). The exchanges and interactions are an essential part of the activity for me. I appreciate bringing people who work on slightly different but connected topics together, for instance by organising events. I also think this is important.
Finally, I like writing: writing to communicate, transfer knowledge and present ideas in a nice way. This is also a very nice activity, even if it can be quite challenging to structure various elements, especially at the start of a new piece of writing.
Basically, meeting people and doing research.
What, in your opinion, are the grand challenges in your field?
I will not answer in terms of grand theoretical challenges, as I think there are more adequate persons to answer in those terms. So if you leave theory aside, I would pick two things.
If you look at the IP field in the broad sense, I would say we still need a bit of unification for some concepts, even if we do already know about many connections. We still need to draw some bridges between concepts that exist or are important for one theory but are a bit disregarded in other theories. Why do I think this is important, beyond being interesting in itself? Firstly, it allows people to see how some ideas from one theory can be interesting to another theory. For example, the notion of inconsistency or incoherence in the information you have. These are important concepts in belief function and possibility theory: how to measure inconsistency and use it in my reasoning process? If you look at lower previsions and connected theories such as desirable gambles, you have some works dealing with inconsistency (from Erik Quaeghebeur, Teddy Seidenfeld, and others), so this is not to say this question is completely disregarded, yet it is usually not regarded as something potentially useful. However, I do think incoherence could have things to say about important problems where incoherence naturally pops ups, such as model choices, logical revision or information fusion (where different experts or sources can disagree with each other).
The other reason I think it is important is about having a common story. If you go to people who do not know about these theories, they might say that you have all these options, and be lost about which one to choose. If you have unification, you can remove this hurdle. When we try to connect to other fields, there is a zoo or jungle, as people say, of uncertainty theories, and you have to find your way around it. This is okay if you have years of PhD in front of you to find your way, not so much if uncertainty is only one of your many problems.
The second aspect that I would consider as important for the field is to really try to connect with other domains. More specifically, how can our modelling tools be helpful in addressing key problems in other domains? By this, I mean not just applying IP to other fields, but searching for problems in other areas currently being solved in perhaps an unsatisfying way, and seeing how our models can explain things in a more satisfying way. This is hard to do, but we probably need to get even more out of the box. It is difficult because you need to engage and get to know other communities; this requires effort as you need to become a listener and get out of your comfort zone, with the possibility of ending up with not much to say. We had some success in doing that, for instance in Quantum Mechanics (notably thanks to Alessio Benavoli), but I strongly think we need to do it more.
What is the most unexpected or surprising thing you discovered?
There’s a nice story here. Ignacio Montes was doing his postdoc in our lab, spending some productive months in Compiègne. It was very enjoyable to have him here. We tried to do something around characterising the extreme points of p-boxes when they were defined over discrete spaces, and we got some results. Among other things that we were looking at, we were looking at the maximum number of points a discrete p-box would have. Ignacio could spend days and days on proofs and examples and counterexamples, patiently looking at extreme points for spaces of 2 elements, 3 elements, up to 4 or 5 elements. Eventually we got the numbers, but we couldn’t see a pattern in these numbers. Searching the sequence of numbers on the net led us to find that they followed the sequence of Pell numbers. We were surprised by this, as none of us knew about Pell numbers before! It really appeared out of the blue. It was a quite surprising discovery, and to this day I have never seen Pell numbers appearing anywhere else.
Lastly, what do you like to spend your time doing outside of work?
Spending time with my family, which I truly enjoy but do not do enough. Reading is also something that would be almost impossible to give up on for me. Walking as this allows me to think about other things. Finally, gaming, both board games with friends and video games.
If you enjoyed this blog, look out for more interview-style blogs in the future.