From electrical networks to neural networks

In May 2014, a team of seven researchers from the 'Systems and Modelling' Research Unit at the University of Liège Faculty of Applied Sciences won the 'Neural Connectomics Challenge: From Imaging to Connectivity'. The aim of this prestigious challenge was to predict the map of connections between neurons in the brain. The team presented its methodology and results during a workshop in September. Interview with the three prize-winners - Damien Ernst, Pierre Geurts and Antonio Sutera - who only found out what a connectome was when they signed up for the competition!

How did your 'Connectomics Challenge' adventure begin?

Pierre: It was Antonio who discovered the competition on a website which listed data analysis competitions (www.kaggle.com). He talked to us about the Connectomics Challenge and we gave him the green light to sign up for it.  Several of our research subjects were close to the issues addressed by this competition. For example, we have  studied gene networks and most of the approaches used in this context can be applied to other types of networks, regardless of their nature.

Damien: The team consisted of five research engineers from the University of Liège's Montefiore Institute in the Department of Electrical Engineering and Computer Science in the Faculty of Applied Sciences. We are a real bunch of friends, some of us specialising in the development of future electrical networks (smart grids) while the others analyse and develop automatic learning algorithms.

Antonio: Arnaud, Vincent, Zixiao, Gilles and I share the same office and this allowed us to compare our ideas and to stimulate our creativity.

We've already heard of genomes and proteomes but what is a connectome?

Damien: It's true that before we signed up for this competition, even we had never heard about connectomes!

Antonio: It is a representation of the structure of the connections between the neurons which make up the brain.

Pierre: The basic structure is very similar from one individual to the next, but we each have our own connectome. In the context of the Connectomics Challenge, we had to work on a group of 1,000 neurons.

What were the various stages of the competition?

Pierre: The aim was to find out how to pair neurons and connect them within the network, based on data which artificially simulated brain activity. We thought we could base our work on a method which we had already designed to study gene networks, but it didn't work. Therefore, we had to develop a new approach.

Antonio: Each neuron has an electrical signal and we measured the correlation between the signals for all neuron pairs.  But we had to be careful, because two neurons may have a similar signal without being connected to one another. We therefore developed an approach which could take into account all the other neurons.

Damien: We received the data in March. It was really exciting because we could follow the developments of the other 143 teams from prestigious universities on-line. We submitted our results in May and were announced as the winners at the end of the competition. We presented our methodology and our results at a workshop in September.(1)

What do you think was your big advantage when it came to winning this competition?

Damien: Compared to the other teams, I think our major advantage was to have a 'naive' approach, focused on effectiveness. Never having ever worked on brain-related issues before, we had no preconceptions before starting the work. We applied the expertise which we had accumulated in other research areas, such as genetic regulation networks and smart grids, to this particular problem.

Antonio: Thanks to the work of the whole team, we could therefore find a relatively simple method which worked well and which we understood. One important aspect for us was to be able to justify each step of our methodology.

Beyond the competition itself, do the organisers have a specific aim?

Damien: The ultimate aim, or the long-term dream, for this kind of research is certainly to be able to one day put a helmet on someone and be able to measure their brain activity and digitally reconstruct the structure of their neural network

Pierre: Practically, it is possible to observe neural activity, but it is impossible to precisely identify the neural network which is behind it. The advantage of using artificial data in the challenge was that the connectome was perfectly known, because it had been defined and the relevant data had been drawn from it.

Damien: Studying in the connectome improves our general understanding of the brain and its learning capabilities. It is also necessary for researching treatments against illnesses which cause alterations to the connectome, such as epilepsy or Alzheimer's disease.

Pierre: Machine learning competitions are relatively new, but they have enabled the research community working in this field to make enormous progress. In this case, the organisers - who are researchers - had developed a tool to identify neural networks based on the data generated and wanted to see if this tool was competitive in comparison with existing techniques, thus motivating other researchers to develop the best tools possible.

How has your participation in this competition influenced your research work, or how will it influence it in the future?

Antonio: Now we have developed this method, we can move towards a more theoretical analysis and to understanding why it worked well. Similarly, we are going to try to theoretically analyse why the method applied to genetics, and which we thought we could use initially in the competition, proved to be ineffective for the connectome. This will enable us to know the limitations of these methods and to identify other systems to which they can be applied.

Damien: Our victory in the competition also opens the way to future collaboration with research teams in neurosciences. Algorithms which have been developed to construct neural networks could certainly be useful for them in terms of applying our method to real data.

Would you be keen to repeat your experience if there were a new edition of this competition?

Damien: For researchers, this type of competition is very interesting. It is interesting from a research point of view, because we find new methods and solutions to problems, but also from a personal point of view because it is fascinating to try to make progress both as a team and as an individual. This also reinforces partnership working within the university and strengthens team spirit. So, the answer is yes, if there was a 2nd edition of this competition, the whole team would be delighted to take part and to accept a new challenge relating to the connectome.

(1)    Antonio Sutera, Arnaud Joly, Vincent François-Lavet, Gilles Louppe, Damien Ernst and Pierre Geurts. Simple connectome inference from partial correlation statistics in calcium imaging. JMLR: Workshop and Conference Proceedings (2014)1–11.  Neural Connectomics Workshop.