I work on statistical learning. It is a multidisciplinary field, between information technology, applied mathematics and statistics. This falls within the scope of what we call data science or artificial intelligence. It concerns transforming data into either scientific knowledge or technological tools. Today, data is everywhere - medical, visual...It concerns physics, the neurosciences, biology...We talk of "big data" and, to process it, scientists need mathematical tools and algorithms.
At Inria, I am working in the Thoth team, which specialises in visual information - be it simple images or video. Our goal is to be able to organise this information automatically and to translate it into words. As far as the ERC project is concerned, the aim is first and foremost to develop very generic tools that will be accessible to scientists in a large number of fields.
With the arrival of big and complex data, current learning techniques must evolve. Major results have been achieved thanks to tools that require the building of a big database that is completely "labelled". In the case of visual data, this signifies manually defining what each part of the image represents. For example, an autonomous car must be able to learn to recognise what a pedestrian is, but also a tree, a road, a road sign...
For this, it requires a large quantity of labelled data. This represents hundreds of hours of videos where we have defined where the road was located, that there might be a pedestrian at a certain junction...however, labelling these enormous amounts of data is very expensive. A major research effort is currently taking place in order to develop methods that are simpler to use and which are capable of exploiting a large quantity of unlabelled data. This is what we call unsupervised learning. Certain labelled data would therefore suffice in order to provide the missing information. Now, in this field humans - even very young children - are infinitely more efficient than machines. It is one of the barriers that we are seeking to remove as part of this ERC project.
During my studies, I quite rapidly turned towards information technology following the preparatory classes, but I was also interested in other fields. What is interesting about statistical learning is the multidisciplinary aspect... For example, at the moment I am working with researchers who are specialised in the processing of genomic data, but also with other experts in the neurosciences. This cross-disciplinary aspect is my main source of motivation.
It is a prestigious honour in the academic world. I had to define and defend a solid and coherent project; this represents a great investment that is prepared a long time in advance. The reflection and writing phase aimed at defining the research project over the long term is particularly interesting. Other aspects are more tedious, such as the communication to "sell" the project to the ERC panel.
Once we have received the grant, we have a period of five years to dedicate ourselves to our research, without having to spend a lot of additional time trying to find funding. In the world of scientific research, five years is a relatively long time. In concrete terms, the money will be used to create a small research team. Thanks to the ERC funding, four or five people will be able to work alongside me full-time.
