Rémi Gribonval: a mathematician of sound and images

Changed on 23/06/2020
This year saw four more young Inria researchers awarded a highly selective grant by the European Research Council (ERC) for the purposes of carrying out exploratory research over 5 years, with a budget of €1 - 1.5 million. One of the winners was Rémi Gribonval, a member of the Metiss project team at the Inria Rennes - Bretagne Atlantique centre. The aim of his project, “Please”, is to develop new methods for signal processing, with applications in the field of audio technology and biomedicine.

The core focus of the Please project is primarily methodological. It aims to explore new methods at the crossroads between signal processing and machine learning, using parsimonious data models. This is fundamental research, but it will also have two main applications. “We will test our methods and models in the field of audio technology and in biomedicine”, explains Rémi Gribonval. “Until now, my work primarily focused on the field of audio technology. “Please” is now looking to innovate by opening up more to biomedicine, with a particular emphasis on medical imaging, drawing on the existing skillsets of Inria personnel in this field. Our research in the field of machine learning could lead to the automatic discovery of biomarkers (specific structures), for example, through the gathering of images of the brain (X-rays, MRI). Identifying them could help when it comes to diagnosing certain neurodegenerative diseases.” Some might consider Rémi Gribonval to be something of a visionary, but in any case, one thing is for sure: his ambitious programme is entirely in keeping with the research into signal processing that he has been carrying out as part of Inria for more than ten years, employing an innovative approach involving data parsimony.

The aim is to develop parsimonious methods in order to concisely describe large collections of sounds and images.

A former student of the École Normale Supérieure in Paris, Rémi Gribonval began by studying for a PhD in mathematical applications in the field of sound at École Polytechnique and the IRCAM (the Institute for Research and Coordination in Acoustics/Music) between 1995 and 1999. He then spent a year at the University of South Carolina in the USA, where he worked on the theoretical aspects of the algorithms that he had developed while studying for his PhD. Upon his return, Rémi Gribonval joined Inria as a postdoctoral researcher, and has remained there ever since.  He began by exploring parsimonious data representation, which has demonstrated its worth in the processing of large volumes of signals and in representing data concisely. “Some algorithms can be used to rebuild data that is missing from an incomplete data set. I demonstrated this theoretically and tested it practically in the field of audio.” Rémi Gribonval's theoretical contributions have proved crucial in popularising the concept of parsimony in signal and image processing. But his research has also focused on other applications. The collaborative project Small, which is funded by the European Union and which Gribonval has been coordinating for two years, recently made a breakthrough in medical imaging, developing compressed sensing techniques for acquiring high resolution images with shortened acquisition times. Applications have also been developed for radar satellite imagery and astronomy. In keeping with this, “‘Please’ will draw on the research carried out by Small, which will come to an end next year”, explains Rémi Gribonval. “The 1.5 million Euro grant that I have just been awarded will enable us to recruit three PhD students, three postdoctoral researchers and to invite researchers to participate in the project”. This was a big achievement for Rémi Gribonval, who is hoping to trigger a new dynamic in his scientific field.

"Please" : parsimony in all its forms : Signals and images constitute extremely large data, and in some cases are comprised of millions of pixels. In the interests of handling them efficiently, the concept of parsimony makes use of the fact that this data can be thought of combinations of a small number of basic shapes (atoms) taken from a dictionary. This traditional, parsimonious approach is similar to building complex-shaped objects using a small number of simple Lego blocks. While parsimony has had a significant impact on signal processing over the past ten years, “the results from the Small project enabled us to identify the potential for new ‘analytical’ models, as opposed to the traditional ‘synthesis’ approach” explains Rémi Gribonval. One component of the Please project involves developing and operating these new models, some of which have highly promising links with equations from physics. The other main component will involve deploying random projection techniques for machine learning. “In medical imaging (MRI) or radio astronomy, random projections have helped reduce acquisition times and to improve image quality. By compressing large collections of data using these tools, Please will provide a more efficient way of extracting the information contained within.