Automatic methods for sparse blind source separation

Christophe Kervazo (LTCI, Télécom Paris, Institut Polytechnique de Paris)
January 27, 2023 — 14:00 — "Salle des séminaire du L2S" (and Teams)

Abstract

Over the last decades, sparse Blind Source Separation (BSS) has become a well-established tool for a wide range of applications. Classical optimization-based sparse BSS methods, such as the Proximal Alternating Linearized Minimization (PALM) algorithm, nevertheless often rely on a cumbersome handcrafted hyper-parameter choice, undermining their practical results and making them difficult to use. In this presentation, we will therefore explore several strategies to bypass this pitfall. We will start by exploring some statistic-based automatic hyper-parameter choice rules, and we will eventually discuss data-driven methods leveraging algorithm unrolling/unfolding. We will furthermore consider an extension of sparse BSS to continuous target extraction in Synthetic Aperture Radar (SAR) images. Overall, we will show that our findings can contribute to a wide range of imaging applications: astrophysics, remote sensing and biomedical imaging, to only name a few.

References

[1] Kervazo, C., Bobin, J., Chenot, C., & Sureau, F. (2020). Use of PALM for ℓ1 sparse matrix factorization: Difficulty and rationalization of a two-step approach. Digital Signal Processing, 97, 102611.

[2] Fahes, M., Kervazo, C., Bobin, J., & Tupin, F. (2021, September). Unrolling PALM for Sparse Semi-Blind Source Separation. In International Conference on Learning Representations.

Biography

Christophe Kervazo received Supélec engineering degree in 2015, and the master of science in Electrical and Computer Engineering from Georgia Institute of Technology (USA) in 2016. From 2016 to 2019, he was a PhD student in the CosmoStat group at CEA Saclay, where he worked on the optimization framework for sparse blind source separation, as well as non-linear component separation. He then went for one year in Mons (Belgium) where he worked, as a post-doctoral researcher, on the extension of Nonnegative Matrix Factorization to Linear-Quadratic mixture unmixing, with mathematical guarantees. He is currently an Assistant Professor at Télécom Paris, in the IMAGES group, were he now mainly works on deep-learning tools for inverse problems, with a focus on algorithm unrolling/unfolding. His works main current application is remote sensing imaging, but he also works on biomedical data.