Material

Abstracts and Presentations

If you would like a copy of any of the technical program presentations, please contact guillaume.ginolhac@univ-smb.fr.

Below are summaries of some of the presentations.

Lectures

Title: Statistical Signal Processing for 4D Imaging Radar-based Autonomous Applications

Speaker: Igal Bilik, Ben Gurion University of the Negev

Abstract: The evolution of autonomous vehicles and drones has shifted radar requirements from simple obstacle detection to high-fidelity environmental perception. Conventional radar signal processing relies on robust statistical frameworks, such as constant false alarm ratio (CFAR) detection, direction-of-arrival (DOA) estimation, and Kalman filtering, to operate in complex, non-Gaussian urban environments. However, as we move toward 4D imaging radar, which provides range, Doppler, azimuth, and elevation, the dimensionality and sparsity of the data pose significant challenges for classical estimators.

This talk explores the paradigm shift from model-based statistical signal processing to data-driven deep learning architectures in the automotive radar domain. We will discuss how the "statistical learning" framework can be applied to solve fundamental radar challenges, including: distributed target detection, statistical radar point cloud semantic segmentation, and achieving high-performance sensing with low-resolution and low-cost hardware, radar-vision sensor fusion, and collaborative sensing, using a network of distributed radars to enhance overall system performance.

By bridging the gap between rigorous statistical theory and the representative power of deep learning, we can achieve the "human-like" perception necessary for the next generation of active safety and fully autonomous navigation. The talk will conclude with a comparative analysis of Radar vs. LiDAR, highlighting why radar's statistical resilience remains the key enabler for reliable, all-weather autonomous sensing. 

Title: Learning to Reconstruct, Learning to Discover

Speaker: Reinhard Heckel, TU München

Abstract: Traditionally, algorithms for signal and information processing are handcrafted. Today, many of the best performing algorithms, especially in tasks like image reconstruction, are learned from data. What is more, machine learning can now assist in discovering novel solutions to long-standing algorithmic challenges. In this talk, I will first present our work on learning efficient algorithms for signal reconstruction, and then discuss how LLM driven agents can aid in finding new solutions to challenging signal and information processing problems. 

Title: Graph Coarsening: The Geometry of Abstraction in Scalable and Trustworthy Graph Learning

Speaker: Sandeep Kumar, The Indian Institute of Technology Delhi

Abstract: Graphs are the language of connected intelligence—capturing relationships across communication, biological, social, and financial systems. As these networks expand in scale, complexity, and sensitivity, learning over them calls for abstraction that is both principled and purposeful. This talk will highlight our group’s recent advances in graph coarsening, reimagining it as the mathematical foundation for scalable, interpretable, and privacy-conscious graph learning. By unifying spectral fidelity, feature alignment, and community preservation within a rigorous optimization framework, we develop algorithms that compress without compromise—retaining the essential geometry and semantics of the original graph while ensuring efficiency, trust, and theoretical soundness. Extending to dynamic, semi-supervised, and neural settings, these developments position graph coarsening as a geometry of abstraction—a bridge between structure and intelligence and a cornerstone for the next generation of scalable, ethical, and adaptive graph machine learning. 

Title: Exact Permutation Recovery under Unknown Scalar Affine Transformations 

Speaker: Arshak Minasyan, CentraleSupélec

Abstract: The problem of matching two sets of features appears in various tasks of computer vision and biomedical applications. The problem can be formalized as a problem of permutation estimation. We address this problem from a statistical point of view when the true feature vectors are connected by an unknown scalar affine transformation. To this end, the minimax rate of separation is investigated and its expression is obtained as a function of the sample size, noise level and dimension of the features. Our result asserts that the obtained rate on the separation distance, under mild heteroscedasticity, coincides with that of the non-affine setting. We additionally demonstrate that there exist configurations requiring a larger minimal separation distance for perfect recovery. The latter makes the matching problem more challenging from a minimax perspective compared to the non-affine setting. Consequently, we show that in the problem of feature matching, standardizing the data implicitly estimates the scalar affine parameters. 

As part of our analysis, we prove non-asymptotic concentration bounds for the affine parameter estimators in the presence of heterogeneous noise magnitudes. 

Short Talks

# Session 1

• Distribution-agnostic linear combining and estimation under heterogeneity, Angelo Coluccia, University of Salento, Italy

Abstract: Determining suitable weights for a linear combiner (aggregator) is a long-standing problem that remains highly relevant across signal processing, data fusion, and machine learning.

While classical estimation approaches have been deeply investigated and perform well in typical settings, more sophisticated (and comparably less explored) weighting strategies are required to handle heterogeneity under challenging conditions such as lack of knowledge about the data distribution and heterogeneous, imbalanced data with low-support or heavy-tailed distribution of sample sizes. This talk discusses different approaches and recent advances in this respect, specifically considering the problem of unbiased linear estimation of a (global) parameter from heterogeneous (local) groups of data, with real-world examples and an outlook on possible future research directions. 

• Covariance Learning for Sparse Signal Recovery, Majdoddin Esfandiari, Univ. Of Aalto

Abstract: tbc

# Session 2

• Deep Learning on Covariance Matrices: Riemannian Optimization, Matthieu Gallet, Univ. Savoie Mont-Blanc, France

Abstract: Symmetric positive definite (SPD) matrices appear naturally in many domains—brain-computer interfaces, hyperspectral imaging, radar—and live on a Riemannian manifold whose geometry must be respected to obtain high-performing methods.We present one contribution around learning on SPD matrices concerning batch normalization in SPDnet, a network whose parameters live on the Stiefel manifold. We formally derive backpropagation via Sylvester equations and propose efficient alternatives to geometric mean that reduce training time by up to a factor of five. A parametric extension, ARMAGNAC, automatically learns the mean adapted to the data.This presentation illustrates how the geometric structure of data induces non-trivial optimization problems, from backpropagation in a neural network onwards. 

• All-pole centroids in the Wasserstein metric with applications to clustering of spectral densities, Rumeshika Pallewela, Univ. Of Aalto

Abstract: In this work, we propose a method for computing centroids, or barycenters, in the spectral Wasserstein-2 metric for sets of power spectral densities, where the barycenters are restricted to belong to the set of all-pole spectra with a certain model order. This may be interpreted as finding an autoregressive representative for sets of second-order stationary Gaussian processes. While Wasserstein, or optimal transport, barycenters have been successfully used earlier in problems of spectral estimation and clustering, the resulting barycenters are non-parametric and the complexity of representing and storing them depends on, e.g., the choice of discretization grid. In contrast, the herein proposed method yields compact, low-dimensional, and interpretable spectral centroids that can be used in downstream tasks. Computing the all-pole centroids corresponds to solving a non-convex optimization problem in the model parameters, and we present a gradient descent scheme for addressing this. Although convergence to a globally optimal point cannot be guaranteed, the sub-optimality of the obtained centroids can be quantified. The proposed method is illustrated on a problem of phoneme classification. (also available on the paper : https://arxiv.org/abs/2602.14583) 

# Session Poster

• G-LaD: Graph-Language alignment for few-shot Diagnosis from fMRI, Abhishek Gupta, IIT Delhi

Abstract: tbc

• Detection and mitigation of interference for satellite systems Alexandre Brochard ENAC, Toulouse

Abstract: tbc

• Fetal Magnetocardiography Reconstruction Using a Biophysical Forward Model Jonas Emrich Technical University of Munich

Abstract: tbc

• MSE Lower Bounds for Single-Tone Signal Estimation and their Use in Array Configuration Optimization Kim Thang Vo SATIE, Paris Saclay, France

Abstract: tbc

# Session 3

• Including Node Textual Metadata in Laplacian-constrained Gaussian Graphical Models, Arnaud Breloy, CNAM, Paris

Abstract: This paper addresses graph learning in Gaussian Graphical Models (GGMs). In this context, data matrices often come with auxiliary metadata (e.g., textual descriptions associated with each node) that is usually ignored in traditional graph estimation processes. To fill this gap, we propose a graph learning approach based on Laplacian-constrained GGMs that jointly leverages the node signals and such metadata. The resulting formulation yields an optimization problem, for which we develop an efficient majorization-minimization (MM) algorithm with closed-form updates at each iteration. Experimental results on a real-world financial dataset demonstrate that the proposed method significantly improves graph clustering performance compared to state-of-the-art approaches that use either signals or metadata alone, thus illustrating the interest of fusing both sources of information. https://arxiv.org/pdf/2602.15920 

• Radar target detection with Support Vector Data description, Jean Pinsolle, SONDRA, CentraleSupelec, France

Abstract: We investigate the use of Support Vector Data Description (SVDD) and its deep extension, Deep SVDD, for target detection. These one-class learning methods avoid direct noise covariance estimation and are adapted here as CFAR detectors. We propose two novel SVDD-based detection algorithms and demonstrate their effectiveness on simulated radar data. 

• Group Variable Selection with False Discovery Rate Control via the T-Rex Selector, Helena Mehler, TU of Darmstadt

Abstract: This talk addresses group variable selection with false discovery rate (FDR) control in high-dimensional linear regression. Building on the T-Rex selector framework, an extension to the group-sparse setting is presented, where predictors are selected at the level of predefined groups rather than individual variables. Group-level sparsity is induced via the iCAP algorithm, a computationally efficient special case of the Composite Absolute Penalties (CAP) family based on combined L1 and L∞ norms. The talk further discusses theoretical guarantees and empirical performance, and concludes with a brief overview of ongoing work. 

# Session 4

• Improving the Posterior Sampling of Plug & Play Diffusion Models by Super-Calibration : application to Poisson inverse problems, Liam Moroy, Heriot Watt University

Abstract: Plug-and-play diffusion models have recently demonstrated strong performance in solving inverse problems. Their key advantage lies in the plug-and-play property, which allows users to modify forward model parameters without retraining the score network. However, applying these models requires specifying an approximation of the time-dependent log-likelihood. To address this, several approximations have been proposed, including Diffusion Posterior Sampling (DPS), pseudo-inverse guidance, MCG, and ILVR. While these methods enable practical use within the plug-and-play framework, they may introduce biases and lead to significant errors in the sampled posterior distribution. In this work, we investigate a new calibration method, that computes an optimal weight to multiply the score of the approximated likelihood. This calibration aims to reduce the discrepancy between the true posterior and its approximation, leading to improved sampling accuracy. 

• Projection-based Riemannian federated learning with partial participation, Thibault Pautrel, Univ. Paris Saclay, France

Abstract: Federated learning on Riemannian manifolds enables collaborative training without centralized data pooling when model parameters are intrinsically constrained. Existing methods either rely on geometric operations lacking closed-form expressions on key manifolds (such as the Stiefel manifold), employ optimizer-specific gradient streams that incur information loss through successive transports, or -- even when computationally cheap -- require drift correction terms and full client participation. We propose two aggregation strategies, RFedProj and RFedRL, that are optimizer-agnostic, lightweight (requiring only standard projections and retractions), and support partial participation without auxiliary correction. Both achieve identical convergence rates under these relaxed conditions and data heterogeneity, with bounds explicitly characterizing how participation ratio and heterogeneity interact -- mirroring classical Euclidean federated guarantees. Experiments on EEG motor imagery classification with SPDNet on compact Stiefel manifolds validate competitive performance against centralized and Euclidean baselines. Then, we will examine the extension of such results to the Gaussian Differential Privacy setting. 

# Session 5

• Exploring Flow Matching for Radar Detection Preeti Meena SONDRA, CentraleSupelec, France

Abstract: Detection in environments with non-Gaussian clutter and noise is a challenging problem. In this talk, I will discuss D-RFM, a method based on Rectified Flow Matching, which maps radar observations to a latent Gaussian space to identify targets as deviations from expected patterns. Unlike traditional detectors that rely on predefined statistical models, D-RFM adapts to the underlying data distribution, capturing complex structures in the observations. I will share insights from preliminary experiments and highlight the effectiveness of this approach for robust radar detection. 

• CGEM: A Context-Aware Expectation Maximization Framework for 2D-3D Registration, Younes Boutiyarzist, Univ. of Toulouse

Abstract: Estimating camera pose by aligning 2D keypoints with a prior 3D map is a common task in visual localization. In sparse mapping scenarios, such as aeronautical navigation, maps often contain only geometry, without texture or appearance descriptors. This makes 2D-3D registration a joint problem of data association and pose estimation. The proposed sparse refinement framework uses learned contextual neighborhood descriptors and is formulated as a probabilistic Expectation--Maximization (EM) method. Correspondence responsibilities combine contextual consistency in an embedding space with reprojection likelihood, instead of relying only on reprojection errors. This improves pose refinement from coarse initial estimates to more accurate localization. 

# Session 6

• A Robust Plug-and-Play Approach for 3D Point Cloud Registration via Expectation-Maximization Maurine BOUZEID Univ. of Toulouse

Abstract: Plug-and-play algorithms have shown impressive results on imaging inverse problems, such as registration, super-resolution, denoising and inpainting. These methods rely on a neural network denoiser to learn an implicit prior of the image to be estimated. This paper investigates a new plug-and-play approach for the registration of two 3D point clouds, when one of these point clouds is corrupted by outliers. The 3D point cloud registration problem is formulated as an inverse problem whose unknowns are the image to be estimated and the transformation between the two point clouds. A plug-and-play approach using an alternating optimization strategy is proposed for solving this inverse problem. Experiments conducted on synthetic data and LiDAR point clouds are presented to evaluate the potential of the proposed method.

• Sound Field Estimation Using Optimal Transport Barycenter in the Presence of Phase Errors, Yuyang Liu, Univ. Of Aalto

Abstract: This study introduces a novel approach for estimating plane-wave coefficients in sound field reconstruction, specifically addressing challenges posed by error-in-variable phase perturbations. Such systematic errors typically arise from sensor mis-calibration, including uncertainties in sensor positions and response characteristics, leading to measurement-induced phase shifts in plane wave coefficients. Traditional methods often result in biased estimates or non-convex solutions. To overcome these issues, we propose an optimal transport (OT) framework. This framework operates on a set of lifted non-negative measures that correspond to observation-dependent shifted coefficients relative to the unperturbed ones. By applying OT, the supports of the measures are transported toward an optimal average in the phase space, effectively morphing them into an indistinguishable state. This optimal average, known as barycenter, is linked to the estimated plane-wave coefficients using the same lifting rule. The framework addresses the ill-posed nature of the problem, due to the large number of plane waves, by adding a constant to the ground cost, ensuring the sparsity of the transport matrix. Convex consistency of the solution is maintained. Simulation results confirm that our proposed method provides more accurate coefficient estimations compared to baseline approaches in scenarios with both additive noise and phase perturbations. 

# Session 7

• Bayesian Learning of Neural Networks using Message Passing, Lily Taylor, Heriot Watt University

Abstract: In this work, we present a generalised message passing framework for training deterministic and stochastic ANNs using approximated Bayesian inference. More precisely, the method allows to approximation the posterior distribution of network parameters. This framework is designed in a modular manner to permit easy adaptation across a range of deterministic and (doubly) stochastic networks. 

• Radar detection using neural networks/VAE/latent space - diffusion network, Alexis Rouzoumka, SONDRA, CentraleSupelec, France

Abstract: We investigate complex-valued Variational AutoEncoders (CVAE) for radar Out-Of-Distribution (OOD) detection in complex radar environments. We proposed several detection metrics: the reconstruction error of CVAE (CVAE-MSE), the latent-based scores (Mahalanobis, Kullback-Leibler divergence (KLD)), and compared their performance against the classical ANMF-Tyler detector (ANMF-FP). The performance of all these detectors is analyzed on synthetic and experimental radar data, showing the advantages and the weaknesses of each detector. 

# Session 8

• Diffusion Models for Graphs: Factual Generation, Edit Sensitivity, and Molecular Language Modeling, Aditya Shahane, IIT Delhi

Abstract: This talk presents a common research direction across two recent works on diffusion models for graph-structured data and their downstream applications. The first part introduces a graph-conditioned diffusion framework for graph-to-sequence generation that addresses two key limitations of standard generation systems: factual grounding and edit sensitivity. The method uses graph–sequence alignment and a graph-aware adaptive noising strategy so that structurally important tokens retain more signal during denoising, leading to outputs that better preserve source facts and respond more predictably to local graph edits. The second part presents a unified diffusion framework for molecule–language modeling, covering both text-conditioned molecule generation and molecule captioning. Here, a token-aware noising schedule preserves chemically salient tokens that are harder to recover, improving both reconstruction fidelity and caption quality. Across both works, the central message is that adaptive, structure-aware diffusion can provide a principled way to improve controllability, faithfulness, and scientific usefulness in generation from structured inputs.

• A Unified Performance Analysis Framework for GNSS Tracking, Emile Ghizzo, ISAE SupAero, Toulouse

Abstract: Signal tracking is a core function in Global Navigation Satellite System (GNSS) receivers, as it directly governs the accuracy, robustness, and continuity of positioning, navigation, and timing solutions. Numerous tracking architectures have been proposed in the literature, including traditional tracking loops, Kalman filter–based and adaptive loop architectures with potential coupling and external aids (e.g., carrier-aiding operation, tight or ultra-tight coupling), as well as new signal designs. However, existing performance models, originally derived for specific architectures and signal types, may not adequately capture the behavior of such diverse tracking configurations. This presentation introduces a new framework for characterizing GNSS tracking performance for general linear architectures and signal models, including coupling and external aids. It first presents the Cramér–Rao bound for GNSS tracking in a general setting, providing a lower bound on performance that jointly depends on the received signal and the dynamic a priori information of the filtering process. It then extends and generalizes classical tracking stress error and mean-square error (MSE) models by incorporating dynamic model uncertainty, prediction errors, and measurement errors from external aids. In addition, a second-order Taylor approximation is derived to compute thermal-noise-induced tracking errors for arbitrary discriminator designs.