Abstract (assisted by AI): This presentation explores the Epoch of Reionization (EoR), a key period in the history of the Universe when the intergalactic medium transitioned from neutral to ionized, using 21 cm hydrogen emission as a probe. Since current radio interferometers face significant challenges in detecting this faint signal due to foreground contamination, instrumental effects, and systematic errors, the presentation introduces scattering transforms as a way to extract information beyond the conventional power spectrum, particularly the non-Gaussian features of the 21 cm signal. It presents the development of a three-dimensional scattering transform tailored to EoR lightcones, designed to capture the signal’s key characteristics and support future foreground separation methods. The results show that this approach can reproduce the statistical properties of simulated lightcones and provide a foundation for separating the cosmological signal from foregrounds. The presentation also introduces a 2+1D scattering transform that combines spatial information with redshift evolution to improve cosmological parameter inference. Using simulations from 21cmFAST, this method achieves tighter parameter constraints than traditional power-spectrum statistics, even in the presence of substantial instrumental noise. Overall, scattering transforms offer a promising framework for analysing 21 cm observations, improving our understanding of the Epoch of Reionization and enabling more informative constraints on the physical processes that shaped the early Universe.

Abstract (assisted by AI): This presentation introduces IViS (Interferometric Visibility-domain Inversion Software), a green and scalable imaging framework designed for the SKA era, developed to process large HI surveys such as GASKAP-HI with ASKAP. It revisits joint deconvolution through a non-linear optimization approach applied directly in the visibility domain, combining multiple beams into a single cost function with spatial regularization to improve image fidelity. Implemented in PyTorch and accelerated with GPU-based NuFFT, IViS aims to address key challenges of deep integration, image quality, and computational sustainability. The method is validated using synthetic observations and real ASKAP data, comparing joint deconvolution with traditional linear mosaicking, and demonstrating its relevance for next-generation large-scale radio surveys.

Abstract: Inverse problems are often ill-posed: incomplete measurements, noise, and ambiguity mean that we must rely on priors to reconstruct meaningful solutions. In the first part of this talk, I will present recent work that analyzes the limitations and challenges of these approaches when solving unsupervised inverse problems—how the choice of priors can bias reconstructions, how conditional models interact with measurement operators, and what this reveals about the fundamental difficulty of filling in missing information. Building on this perspective, I will then introduce FIRE (Fixed point Restoration), a framework that addresses these challenges by defining implicit priors not just through denoisers but through general restoration models. The key idea is to characterize natural signals as fixed points of a degradation–restoration cycle, enabling a principled and flexible way to integrate pretrained networks into inverse problem solvers. This fixed-point view not only broadens the class of usable priors but also leads to robust, algorithms with strong empirical performance.

Resume (assisted by AI): This presentation is about acoustic source localization and imaging with microphone arrays, covering both theory and advanced algorithms. It introduces the problem and classical source-localization applications, then explains beamforming—highlighting steering-vector choices, bias/variance behavior, and limits in resolution. It then presents DAMAS and Covariance Matrix Fitting (CMF), clarifying convergence properties and offering faster implementations, along with comparisons to CLEAN-type methods. The talk also explores gridless approaches such as BLASSO and Sliding Frank–Wolfe for super-resolution imaging, as well as extensions to correlated sources with greedy algorithms and asynchronous arrays analyzed using maximum-likelihood estimation and Cramér–Rao bounds. Throughout, simulated and experimental results (2D and 3D) demonstrate performance, robustness at low SNR, and computational considerations.

Resume (assisted by AI): This work explores how antenna array configurations influence image reconstruction performance in radio interferometry and proposes using estimation error lower bounds as design criteria. After presenting the interferometric measurement model, uv-plane sampling, and point-spread function formation, the study examines classical and advanced MSE bounds—such as the Cramér–Rao and Barankin bounds—and illustrates their relevance through a frequency estimation example. These bounds are then applied to a single-source direction-of-arrival model to guide optimal antenna placement, demonstrating improved performance over traditional criteria. The approach highlights both the potential of statistically grounded design and its current limitations, motivating ongoing efforts within the ANR PerDRIX project to develop more appropriate cost functions, scalable optimization methods, and more realistic interferometric models.

Resume (assisted by AI): This presentation exposes an overview of the science of neutral hydrogen (HI) in galaxies, focusing on 21 cm observations and their role in understanding galactic structure, dynamics, and star formation processes. They highlight methodological advances in hyperspectral data analysis, such as Gaussian decomposition and CNM/WNM separation, which enable deeper insights into the diffuse interstellar medium. The GASKAP project, a major ASKAP key survey, is introduced as the largest HI mapping effort to date, covering the Milky Way and Magellanic System with unprecedented resolution and data volume. Processing challenges—such as handling petabyte-scale datasets, imaging diffuse emission across multiple spatial scales, and developing new joint deconvolution techniques—are discussed as crucial steps toward preparing for the SKA. Ultimately, GASKAP is presented as a transformative resource for studying HI from sub-parsec to 100 kpc scales, bridging local and extragalactic environments.

Resume (assisted by AI): This presentation explores the study of the Cosmic Dawn and the Epoch of Reionization (EoR) through 21-cm line observations using LOFAR and NenuFAR radio telescopes. It highlights the challenges of detecting the faint 21-cm signal, which is buried under strong astrophysical foregrounds and instrumental effects, and presents various mitigation strategies, including sky model optimization, point-source subtraction, and advanced machine learning techniques such as Gaussian Process Regression (GPR) and Variational Autoencoders. The LOFAR-EoR project has established new multi-redshift upper limits on the 21-cm power spectrum, rejecting cold IGM scenarios, while NenuFAR achieved the deepest upper limits at z ~ 20 using newly identified dark sky regions. The presentation culminates with results from the SKA Data Challenge 3a, where the DOTSS-21 team demonstrated a complete pipeline for extracting the 21-cm signal by combining compact source subtraction, galactic foreground modeling, and ML-GPR, setting the stage for future breakthroughs in 21-cm cosmology.

Resume (assisted by AI): The presentation by Nicolas Monnier, focuses on optimizing computational costs in radio interferometric imaging by introducing the Grid to Grid (G2G) method. After introducing the Radio Interferometric Measurement Equation (RIME) and challenges of discretization, especially for large arrays like SKA, Monnier explains that iterative imaging algorithms are bottlenecked by costly gridding and degridding steps. The G2G approach efficiently fuses these operations, drastically reducing memory use and computational demands, achieving up to 10× speedup over traditional methods when implemented on GPUs, while maintaining accuracy. Experimental results on VLA datasets confirm the method’s strong performance and reduced memory footprint compared to standard approaches.

Resume (assisted by AI): This presentation explores efficient learned deconvolution of radio interferometric images using deep unrolling techniques. With the construction of large observatories like SKA generating massive datasets, traditional image reconstruction methods face computational challenges. The presentation introduces algorithm unrolling, which transforms iterative reconstruction methods into deep neural networks, significantly improving speed and performance. It also discusses the CARB (Conformalized Augmented Radio Bootstrap) method for uncertainty quantification, providing reliable confidence intervals without requiring ground truth data. The results demonstrate that unrolling achieves high-quality reconstructions with reduced computational costs, making it a promising approach for large-scale radio astronomy imaging.