ISP Seminars

Previous Seminars

Exact Continuous Relaxations of $L_0$-Regularized Generalized Linear Models
Emmanuel Soubies (IRIT, CNRS, Université de Toulouse)
22 August 2024
Sparse generalized linear models are widely used in fields such as statistics, computer vision, signal/image processing and machine learning. The natural sparsity promoting regularizer is the $L_0$ pseudo-norm which is discontinuous and non-convex. In this talk, we will present the $L_0$-Bregman relaxation (B-Rex), a general framework to compute exact continuous relaxations of such $L_0$-regularized criteria. Although in general still non-convex, these continuous relaxations are qualified as exact in the sense that they let unchanged the set of global minimizer while enjoying a better optimization landscape. In particular, we will show that some local minimizers of the initial functional are eliminated by these relaxations. Finally, these properties will be illustrated on both sparse Kullback-Leibler regression and sparse logistic regression problems. This is joint work with M'hamed Essafri and Luca Calatroni.

External Web page: https://www.irit.fr/~Emmanuel.Soubies/

Euler seminar room (a.002) -- Thursday, 22 August 2024 at 14:00 (50 min.)

Computational Imaging: Restoration Deep Networks as Implicit Priors
Ulugbek Kamilov (Washington University, St. Louis, USA)
28 May 2024
Many interesting computational imaging problems can be formulated as imaging inverse problems. Since these problems are often ill-posed, one needs to integrate all the available prior knowledge for obtaining high-quality solutions. This talk will explore a series of techniques that leverage deep neural networks for image restoration as data-driven, implicit priors for images. The methods discussed originate from the well-known plug-and-play (PnP) methodology, known for its effectiveness in addressing imaging inverse problems. We will extend the conversation to the generalization of PnP methods, moving beyond traditional use of additive white Gaussian noise (AWGN) denoisers to include a variety of other restoration networks. This expansion not only enhances imaging performance but also offers the flexibility to train priors in the absence of clean data. Additionally, the talk will cover the theoretical underpinnings of using deep restoration networks and their applications in biomedical image reconstruction.

Biography: Ulugbek Kamilov is the Director of Computational Imaging Group and an Associate Professor of Electrical & Systems Engineering and Computer Science & Engineering at Washington University in St. Louis. He is currently a Visiting Professor at the Data Science Center at École Normale Supérieure in Paris. He obtained the BSc/MSc degree in Communication Systems in 2011 and the PhD degree in Electrical Engineering in 2015 from EPFL. He was a Visiting Research Faculty at Google Research in 2023-2024 and a Research Scientist at Mitsubishi Electric Research Laboratories in 2015-2017. He was an Exchange Student at Carnegie Mello University in 2008, Visiting Student Researcher at MIT in 2011, and Visiting Scholar at Stanford University in 2013.

He is a recipient of the NSF CAREER Award and the IEEE Signal Processing Society’s 2017 Best Paper Award. He was among 55 early-career researchers in the USA selected as a Fellow for the Scialog initiative on “Advancing Bioimaging” in 2021. His PhD thesis was selected as a finalist for the EPFL Doctorate Award in 2016. He was awarded Outstanding Teaching Award from the Department of Electrical & Systems Engineering at WashU in 2023. He is currently a Senior Member of the Editorial Board of IEEE Signal Processing Magazine and is on IEEE Signal Processing Society’s Bioimaging and Signal Processing Technical Committee. He has previously servedas an Associate Editor of IEEE Transactions on Computational Imaging and on IEEE Signal Processing Society’s Computational Imaging Technical Committee.

External Web page: https://engineering.wustl.edu/faculty/Ulugbek-Kamilov.html

Shannon seminar room (Maxwell) -- Tuesday, 28 May 2024 at 14:00 (45 min.)

AI in computational imaging: from algorithms to radio astronomy
Yves Wiaux (Heriot-Watt University Edinburgh, UK)
23 February 2024
Endowing advanced imaging instruments such as telescopes and medical scanners with an acute vision that enables them to probe the Universe or human body with precision is a complex mathematical endeavour. It requires solving challenging inverse problems for image formation from observed data. In this talk, we will dive into this field of computational imaging, and its specific application in radio astronomy, where algorithms are currently facing a multi-faceted challenge for the robust reconstruction of images at extreme resolution and dynamic range, and from extreme data volumes. We will discuss advanced algorithms at the interface of optimisation and deep learning theories, from SARA, an optimisation algorithm propelled by handcrafted regularisation priors, to AIRI, plug-and-play algorithm relying on learned regularisation denoisers, and the newborn deep neural network series R2D2, which can be interpreted as a learned version of the Matching Pursuit algorithm. If time allows, we will also briefly illustrate the transfer of such algorithms to medical imaging. Last but not least, we will take a few seconds to unveil Star Wars hidden facts and misconceptions.

Biography: Yves Wiaux received the MSc degree in Physics and the PhD degree in Theoretical Physics from the Université catholique de Louvain (UCL, Louvain-la-Neuve) in Belgium, in 1999 and 2002 respectively. He was a Senior Researcher at the Signal Processing Laboratories of the Ecole Polytechnique Fédérale de Lausanne (EPFL) in Switzerland from 2003 to 2013, where he created the Biomedical and Astronomical Signal Processing (BASP) group. In 2013, he moved as an Associate Professor at the School of Engineering and Physical Sciences of Heriot-Watt University where he currently runs the BASP group. He was promoted to Professor at Heriot-Watt in 2016. He is also an Academic Guest at EPFL and a Honorary Fellow at the University of Edinburgh (UoE). Among other responsibilities, Prof. Wiaux chairs the “BASP Frontiers” Conference series, and is an Associate Editor of the IEEE “Transactions in Computational Imaging” (TCI) journal and of the “Royal Astronomy Society Techniques and Instruments” (RASTI) journal. Since 2010, he has led numerous research projects funded by both the Swiss National Science Foundation (SNSF) and the UK Research and Innovation Councils (UKRI). The ethos of Prof. Wiaux’s BASP group is to develop cutting-edge research in computational imaging, from theory and algorithms to applications in astronomy and medicine.

External Web page: https://researchportal.hw.ac.uk/en/persons/yves-wiaux

Euler Seminar Room (A002) -- Friday, 23 February 2024 at 11:00 (45 min.)

V-DMC : the emerging MPEG standard for dynamic mesh compression
Patrice Rondao, Nokia Bell Labs
7 December 2023
MPEG is well-known for its success story in delivering video and audio compression standards that we all use every day, perhaps without noticing it, in our mobile phones, TV, cameras, computers… However, did you know MPEG also produced 3D and immersive representations compression standards? In this talk, I will first shortly explain why these 3D representations are important for a growing number of applications and devices and how MPEG has addressed their compression in a family of new standards called V3C. I will then focus on the ongoing standardization efforts related to the newest member of this family; the emerging V-DMC standard that is designed to efficiently compress 3D dynamic mesh sequences. We will also see how it compares with the relevant state of the art.

Euler seminar room, a.002 -- Thursday, 7 December 2023 at 14:30 (45 min.)

"Screening" and beyond: A strategy to deal with large-scale optimization problems
Cédric Herzet (INRIA/IRMAR - UMR 6625, France)
23 August 2023
In my presentation, I will develop the concept of « safe screening », a technique to reduce the dimensionality of large-scale optimization problems that exploits the sparsity of the solutions. I will first explain the fundamental ingredients of screening and highlight some recent advances in this field. In a second part, I will present how the concept of screening can be extended by exploiting partial knowledge about the solution of the optimization problem.

External Web page: https://people.rennes.inria.fr/Cedric.Herzet/Cedric.Herzet/Main.html

Euler Seminar Room (A002) -- Wednesday, 23 August 2023 at 11:00 (45 min.)

How to Train Better: Exploiting the Separability of Deep Neural Networks
Elizabeth Newman (Emory University, USA)
9 March 2023
Deep neural networks (DNNs) have gained undeniable success as high-dimensional function approximators in countless applications. However, there is a significant hidden cost behind the success - the cost of training. Typically, DNN training is posed as a stochastic optimization problem with respect to the learnable DNN weights. With millions of weights, a non-convex and non-smooth objective function, and many hyperparameters to tune, solving the training problem well is no easy task. In this talk, we will make DNN training easier by exploiting the separability of common DNN architectures; that is, the weights of the final layer of the DNN are applied linearly. We will leverage this linearity in two ways. First, we will approximate the stochastic optimization problem deterministically via a sample average approximation. In this setting, we can eliminate the linear weights through variable projection (i.e., partial optimization). Second, in the stochastic optimization setting, we will consider a powerful iterative sampling approach to update the linear weights, which notably incorporates automatic regularization parameter selection methods. Throughout the talk, we will demonstrate the efficacy of these two approaches through numerical examples.

External Web page: http://math.emory.edu/~enewma5/

Euler a.002 -- Thursday, 9 March 2023 at 14:00 (50 min.)

Diffusion models : from DDPM to DALLE-2
Mathieu Simon (UCLouvain)
22 July 2022
Many of us have probably been mind blowned by the recent progress in text-to-image generation tasks demonstrated by models such as GLIDE, DALLE-2 or Imagen. A lot of these incredible results can be attributed to the meteoric rise of diffusion models, a type of generative model that has been gaining significant popularity recently and has shown to outperform GANs on image synthesis. This presentation aims at giving a small (and humble) introduction to these models and to share my understanding on how we went from the original denoising diffusion probabilistic model to DALLE-2.

External References

Ressources :

Nyquist seminar room, ICTEAM -- Friday, 22 July 2022 at 00:00 (50 min.)

How to facilitate collaboration between pathologists and engineers in AI research applied to Digital Pathology using Cytomine
Grégoire Vincke (Cytomine Corporation SA, Liège, Belgium) (Invited talk)
15 September 2021
Cytomine is a full-web open-source platform dedicated to the collaborative analysis of Whole-Slide Images (WSI) and the development of AI-driven analysis solutions adapted to Digital Pathology. Cytomine can be used in teaching, research, or diagnosis, but in this seminar, we will focus on the features dedicated to the collaboration between pathologists and AI engineers. Cytomine fosters collaboration between pathologists and AI engineers by providing working interfaces adapted to each community. Using our user-friendly web interface, pathologists are able to manage and explore the WSI and enrich them with semantic added value like annotations or terms. Using our java or python external clients, AI engineers are able to use all these images, annotations, and information to develop AI algorithms adapted to Digital Pathology, and train or run them in any computing center. All the AI-generated information is then available inside the web interface to allow pathologists to validate or edit it, using dedicated web tools. If needed, AI apps can be run by end-users directly from inside the web interface. In this seminar, we will present all this workflow alternatively from the pathologists' and the AI engineers' point of view, and by emphasizing the development possibilities offered by Cytomine open source license.

Shannon seminar room -- Wednesday, 15 September 2021 at 14:00 (45 min.)

How to facilitate collaboration between pathologists and engineers in AI research applied to Digital Pathology using Cytomine
Grégoire Vincke (Cytomine Corporation SA, Liège, Belgium)
15 September 2021
Cytomine is a full-web open-source platform dedicated to the collaborative analysis of Whole-Slide Images (WSI) and the development of AI-driven analysis solutions adapted to Digital Pathology. Cytomine can be used in teaching, research, or diagnosis, but in this seminar, we will focus on the features dedicated to the collaboration between pathologists and AI engineers. Cytomine fosters collaboration between pathologists and AI engineers by providing working interfaces adapted to each community. Using our user-friendly web interface, pathologists are able to manage and explore the WSI and enrich them with semantic added value like annotations or terms. Using our java or python external clients, AI engineers are able to use all these images, annotations, and information to develop AI algorithms adapted to Digital Pathology, and train or run them in any computing center. All the AI-generated information is then available inside the web interface to allow pathologists to validate or edit it, using dedicated web tools. If needed, AI apps can be run by end-users directly from inside the web interface. In this seminar, we will present all this workflow alternatively from the pathologists' and the AI engineers' point of view, and by emphasizing the development possibilities offered by Cytomine open source license.

External Web page: https://cytomine.com

Shannon seminar room -- Wednesday, 15 September 2021 at 14:00 (45 min.)

Computational Imaging with Plug-and-play priors: Leverage the Power of Deep Learning
Xiaojian Xu (Washington University in St. Louis)(Invited talk)
9 December 2020
Plug-and-play priors (PnP) is a methodology for regularized image reconstruction that specifies the prior through an image denoiser. Unlike traditional regularized inversion, PnP does not require the prior to be expressible in the form of a regularization function. This flexibility, on one hand, enables PnP algorithms to exploit the most effective image denoisers, such as the ones based on convolutional neural networks (CNNs), leading to their state-of-the-art performance in various imaging tasks. While on the other hand, it also makes it hard to establish the theoretical analysis on the convergence for an arbitrary denoiser. So in this talk, we will show some recent achievement about PnP regarding these issues. We will introduce a denoiser scaling technique, which in practice could greatly enhance the performance and flexibility of PnP for various denoisers, especially those CNN-based ones. We will also introduce the first theoretical convergence result for the iterative shrinkage/thresholding algorithm (ISTA) variant of PnP for MMSE denoisers, where iterates produced by PnP-ISTA with an MMSE denoiser converge to a stationary point of some global cost function.
Ressources :

(Online) Microsoft Teams -- Wednesday, 9 December 2020 at 16:00 (45 min.)