Lecturers
Title: Proximal Neural Networks
Bio: : Prof. Pustelnik completed her Ph.D. at Université Paris-Est in France in 2010. Afterward, she held a post-doctoral position in collaboration with Bordeaux University and TOTALEnergies SA. In 2011, she became a permanent CNRS researcher at the Physics Lab in Ecole Normale Superieur de Lyon, France. From 2019 to 2022, Prof. Pustelnik joined the Applied Mathematics department (INMA) at UCLouvain to further develop her expertise and foster European collaboration in inverse problems and optimization.
Prof. Pustelnik delved into the theoretical and practical aspects of inverse problems solving for signal and image processing, investigating advanced multiresolution techniques and algorithms. Her current research activity focuses on inverse problem solving and physics informed neural networks for observational sciences (astronomy, oceanography, rivers dynamics at the global scale). Her goal is to develop architectures that leverage both domain-specific knowledge and advanced computational techniques to propose stable and energy-efficient methods for solving inverse problems.
In addition to her academic positions, Prof. Pustelnik has actively contributed to the scientific community through various roles in scientific management. She has been recognized for her expertise and contributions by the EURASIP Signal and Data Analytics for Machine Learning SAT and the IEEE MLSP TC, where she actively contributed to the development and dissemination of knowledge in machine learning and signal processing. Additionally, she holds the position of Associate Editor and the Senior Associate Editor for IEEE Signal Processing Letters and she also serves as an Associate Editor for the IEEE Transactions on Image Processing. She has been recently elected in the EURASIP Board of Directors.
Throughout her career, Prof. Pustelnik has actively participated in the organization of prestigious conferences and workshops, including EUSIPCO, STATOS, IEEE ICIP, IEEE IVMSP, and IEEE MLSP by assuming various responsibilities such as student chair, program chair, tutorial chair, and more recently financial chair for EUSIPCO 2024. Prof. N. Pustelnik's research, along with her active participation in the academic signal and image processing community, aspires to further mold the future and the visibility of the discipline.
Abstract: Since the early 2000s, signal and image processing has been profoundly influenced by two major trends: sparsity-driven proximal algorithms and deep learning. The former capitalize on the clever integration of variational frameworks and optimization techniques, while the latter relies on complex neural network architectures.
Both approaches have achieved impressive results in a wide range of applications, with deep learning often outperforming purely optimization-based methods in practical contexts. Nevertheless, for many decision-making tasks, optimization methods are still preferred due to their strong theoretical guarantees for reliable solutions.
More recently, there has been a growing interest in hybrid approaches that combine optimization and deep learning, delivering performance at least on par with traditional deep learning, while offering theoretical guarantees and improved interpretability.
As both proximal algorithms and deep learning have now reached advanced levels of maturity and complexity, there is a valuable opportunity to explore the synergies between these two methodological approaches.
Title: Wider Horizons in Optical Imaging with the Mesolens
Bio: Gail McConnell is Professor of Biophotonics at the Strathclyde Institute of Pharmacy and Biomedical Sciences at the University of Strathclyde, Glasgow, UK. Following a first degree in Laser Physics and Optoelectronics (1998) and PhD in Physics from the University of Strathclyde (2002), she obtained a Personal Research Fellowship from the Royal Society of Edinburgh (2003) and a Research Councils UK Academic Fellowship (2005), securing a readership in 2008 and professorship in 2012. The work in Gail’s multidisciplinary research group involves the design, development and application of linear and nonlinear optical instrumentation and new methods for biomedical imaging, from the nanoscale to the whole organism. She is a Fellow of the Royal Society of Edinburgh, a Fellow of the Institute of Physics, and a Fellow of the Royal Microscopical Society, where she is the current Vice Chair of the Light Microscopy Committee.
Abstract: Since the invention of the microscope, optics have been optimized to match the performance of the human eye. However, since the advent of sensitive and advanced photodetectors the human eye is no longer a limitation. This has opened up exciting possibilities for new instrumentation in biomedical imaging. We have developed an objective we call the Mesolens that can study large unusually large objects with sub-cellular resolution. The pupil size of the lens is so great that it cannot be used with a conventional microscope frame, so we have built the imaging system around the giant lens. Like the original optical microscope, we have found that the Mesolens has a wide range of applications in biomedical research. I will present an overview of the Mesolens imaging technology, and I will show how we are using it to reveal new information from large biological and clinical specimens.
Matthieu Terris (INRIA, France)
Samuel Hurault (CNRS, France)
Title: DeepInverse: A Pytorch Library For Solving Imaging Inverse Problems With Deep Learning.
Bios: Julián Tachella received the electronic engineering degree (Hons.) from Instituto Tecnológico de Buenos Aires, Argentina, in 2016, and the Ph.D. degree from Heriot-Watt University, U.K., in 2020. He held a postdoctoral position at the University of Edinburgh, U.K., from 2020 to 2021. He currently holds a Centre National de Recherche Scientifique (CNRS) researcher position at the École Normale Supérieure de Lyon, France. His research lies at the intersection of signal processing and machine learning. He is particularly interested in the theory of imaging inverse problems and applications in computational imaging.
Matthieu Terris is a postdoctoral researcher at INRIA, focusing on machine learning techniques inverse problems. He earned his PhD from the Biomedical and Astronomical Signal Processing laboratory at Heriot-Watt University. His research interests include scalable computational imaging algorithms relying on learned priors, particularly in the context of radio astronomy and biomedical imaging.
Samuel Hurault is a CNRS postdoctoral researcher at ENS Paris under the supervision of Gabriel Peyré. He studied at ENS Cachan and received its PhD from University of Bordeaux in 2023. His main research interests include deep methods for image inverse problems, convex and nonconvex optimization and generative modeling.
Abstract: DeepInverse https://deepinv.github.io/ is an open-source pytorch library for solving imaging inverse problems with deep learning, which covers most of the steps in modern imaging pipelines, from the definition of the forward sensing operator to the definition of reconstruction algorithms with deep networks. This tutorial will provide a gentle introduction into the library, covering most key reconstruction approaches, such as plug-and-play methods, diffusion algorithms, unrolled architectures, self-supervised learning losses and more. The tutorial will be supported by the core developing team of deepinverse, which will help to get you started with the library in interactive coding sessions.
