Julien Mairal from Inria has received the 2013 ERCIM Cor Baayen Award. Honorary Mentions are given to Alexander Bertrand, KU Leuven and Ben Glocker, Microsoft Research Cambridge. The Cor Baayen Award is awarded annually by ERCIM to a promising young researcher in the field of Informatics and Applied Mathematics.
Julien Mairal obtained his PhD in 2010 at Inria under the joint supervision of Francis Bach and Jean Ponce. His PhD thesis was concerned with new and efficient algorithms for sparse coding and dictionary learning with applications to image processing and computer vision.
Julien’s most significant contribution is a stochastic optimization algorithm for a problem know as dictionary learning in signal processing and machine learning.Given a dataset represented by high-dimensional vectors, the task consists of approximating each data point by a linear combination of a few elements from a learned dictionary. The original dictionary learning formulation, which was introduced a decade ago by Olshausen and Field from the neuroscience community, has recently gained a large interest in signal processing and machine learning, showing promising results for numerous applications. The proposed algorithm was shown to be much more scalable and computationally efficient than existing approaches, which is of acute importance currently with the emergence of massive datasets and the need to analyze them. This work is available through anopen-source software, which is now widely used.
Julien also developed state-of-the-art image restoration algorithms for image denoising and image demosaicking (reconstructing colour images from raw data from CCD sensors), by combining two successful paradigms: sparse representations of image patches, and non-local image self-similarities. His third contribution was to develop new formulations of dictionary learning for classification or regression tasks, whereas traditional applications were devoted to signal or image reconstruction. This proved particularly useful in computer vision, where it allows discriminative representations of natural image patches to be found.
Finally, Julien and his collaborators worked on solving optimization problems related to structured sparse estimation, which has received recently a lot of attention in statistics and machine learning. Structured sparse models differ from classical ones in that variables are grouped together according to an a priori structure, a hierarchy for example. Computing penalized likelihood estimators is challenging in this context, but Julien and his collegues obtained efficient and scalable optimization techniques by drawing novel connections between network flows and sparse estimation. This allows the use of structured sparse models at a large scale, which was not previously possible.
Julien’s PhD work was published in top conferences (CVPR, ECCV, ICCV, ICML, NIPS) and journals (JMLR, PAMI) of computer vision and machine learning, and was honoured by three French prizes for a best PhD thesis. It is also widely cited, with ten papers having received over 100 citations according to Google Scholar, and an amazing h-index of 19 for such a young researcher.
2013 Cor Baayen Award
Winner: Julien Mairal
Honorary Mentions: Alexander Bertrand, Ben Glocker
Finalists: Igor Boehm, André Chailloux, Marek Cygan, Jose Luis Fernandez-Marquez, Gerard Hoekstra, Mustafa Misir Anna Monreale, Stefanie Nowak, Mathias Payer, Sini Ruohomaa, Floor Sietsma, Pavel Vácha, Rossano Venturini.