Publications

2025

• Soft Disentanglement in Frequency Bands for Neural Audio Codecs
  
  • Giniès Benoît
  • Bie Xiaoyu
  • Fercoq Olivier
  • Richard Gaël
  , 2025. In neural-based audio feature extraction, ensuring that representations capture disentangled information is crucial for model interpretability. However, existing disentanglement methods often rely on assumptions that are highly dependent on data characteristics or specific tasks. In this work, we introduce a generalizable approach for learning disentangled features within a neural architecture. Our method applies spectral decomposition to time-domain signals, followed by a multibranch audio codec that operates on the decomposed components. Empirical evaluations demonstrate that our approach achieves better reconstruction and perceptual performance compared to a state-of-the-art baseline while also offering potential advantages for inpainting tasks.
• When Can Sequence Modelling Approaches Recover the Target Policy In Offline Reinforcement Learning? a Statistical Analysis
  
  • Ghani Abdelghanem
  • Ciblat Philippe
  • Ghogho Mounir
  , 2025. <div><p>We present a theoretical analysis of sample complexity for learning the target policy in offline reinforcement learning (RL) using sequence modeling approaches. Our main theorem establishes bounds on the minimum required number of high-return samples. We identify distinct small-data and largedata regimes, characterized by a critical transition point, and reveal a potential trade-off between context coverage breadth and sampling depth. These findings offer insights into efficient data collection strategies and algorithm design for offline RL.</p></div>
• Bayesian Experimental Design with Mutual Information and Learned Errors for Human-Computer Interaction
  
  • Miquel Hugo
  • Gori Julien
  • Rioul Olivier
  , 2025. This work provides a Bayesian framework for handling user errors in interactive systems, with applications in human-computer interaction (HCI) and user modeling. The Bayesian Information Gain (BIG) algorithm [1, 2, 3, 4] is an iterative variant of Bayesian experimental design with mutual information as a cost function, used in HCI. It is a principled approach that maximizes expected information gained from each interaction. More precisely, let Θ be the potential target in the user’s mind with prior distribution p(θ), X be the system feedback, and Y be the corresponding user’s input. In each interaction loop, BIG selects feedback x that maximizes mutual information I(Θ; Y|X= x), assuming a known user model (likelihood) p(y|x,θ), and then updates the posterior distribution p(θ|x,y). This work extends the BIG algorithm to learn from user errors while preserving its mathematical foundations. We incorporate an error rate parameter ϵinto the likelihood function p(y|x,θ,ϵ) and develop an adaptive algorithm that jointly infers both θ and ϵ by updating the posterior p(θ,ϵ|x,y) at each interaction step. We also discuss three simplifying hypotheses for the prior expression p(θ,ϵ) and three user models: (i) zero error; (ii) fixed error rate; (iii) arbitrary random error rate. We prove mathematical continuity between these three models, showing that our adaptive approach naturally extends BIG. We also investigate model mismatch on the overall performance and degradation properties with respect to the standard BIG algorithm. While standard BIG converges quickly with perfect responses, it degrades with even small error rates. The fixed-error model depends critically on correctly estimating the error parameter, while our adaptive model achieves the highest accuracy under varying error conditions, at the expense of additional interactions.
• Causal decompositions of one-dimensional quantum cellular automata
  
  • Vanrietvelde Augustin
  • Mestoudjian Octave
  • Arrighi Pablo
  , 2025. Understanding quantum theory's causal structure stands out as a major matter, since it radically departs from classical notions of causality. We present advances in the research program of causal decompositions, which investigates the existence of an equivalence between the causal and the compositional structures of unitary channels. Our results concern one-dimensional Quantum Cellular Automata (1D QCAs), i.e. unitary channels over a line of N quantum systems (with or without periodic boundary conditions) that feature a causality radius r: a given input cannot causally influence outputs at a distance more than r. We prove that, for N ≥ 4r +1, 1D QCAs all admit causal decompositions: a unitary channel is a 1D QCA if and only if it can be decomposed into a unitary routed circuit of nearest-neighbour interactions, in which its causal structure is compositionally obvious. This provides the first constructive form of 1D QCAs with causality radius one or more, fully elucidating their structure. In addition, we show that this decomposition can be taken to be translation-invariant for the case of translation-invariant QCAs. Our proof of these results makes use of innovative algebraic techniques, leveraging a new framework for capturing partitions into non-factor sub-C* algebras.
• Audio processor parameters: estimating distributions instead of deterministic values
  
  • Peladeau Côme
  • Fourer Dominique
  • Peeters Geoffroy
  , 2025, pp.275-282. Audio effects and sound synthesizers are widely used processors in popular music. Their parameters control the quality of the output sound. Multiple combinations of parameters can lead to the same sound. While recent approaches have been proposed to estimate these parameters given only the output sound, those are deterministic, i.e. they only estimate a single solution among the many possible parameter configurations. In this work, we propose to model the parameters as probability distributions instead of deterministic values. To learn the distributions, we optimize two objectives: (1) we minimize the reconstruction error between the ground truth output sound and the one generated using the estimated parameters, as is it usually done, but also (2) we maximize the parameter diversity, using entropy. We evaluate our approach through two numerical audio experiments to show its effectiveness. These results show how our approach effectively outputs multiple combinations of parameters to match one sound.
• Partitions in quantum theory
  
  • Vanrietvelde Augustin
  • Mestoudjian Octave
  • Arrighi Pablo
  , 2025. Decompositional theories describe the ways in which a global physical system can be split into subsystems, facilitating the study of how different possible partitions of a same system interplay, e.g. in terms of inclusions or signalling. In quantum theory, subsystems are usually framed as sub-C* algebras of the algebra of operators on the global system. However, most decompositional approaches have so far restricted their scope to the case of systems corresponding to factor algebras. We argue that this is a mistake: one should cater for the possibility for non-factor subsystems, arising for instance from symmetry considerations. Building on simple examples, we motivate and present a definition of partitions into an arbitrary number of parts, each of which is a possibly non-factor sub-C* algebra. We discuss its physical interpretation and study its properties, in particular with regards to the structure of algebras' centres. We prove that partitions, defined at the C*-algebraic level, can be represented in terms of a splitting of Hilbert spaces, using the framework of routed quantum circuits. For some partitions, however, such a representation necessarily retains a residual pseudo-nonlocality. We provide an example of this behaviour, given by the partition of a fermionic system into local modes.
• Semi-supervised graph learning for underwater source localization using ship-of-opportunity spectrograms
  
  • Castro-Correa Jhon
  • Badiey Mohsen
  • Giraldo Jhony
  • Malliaros Fragkiskos
  Journal of the Acoustical Society of America, Acoustical Society of America, 2025, 158 (3), pp.1836-1848. Conventional techniques for underwater source localization have traditionally relied on optimization methods, matched-field processing, beamforming, and, more recently, deep learning. However, these methods often fall short to fully exploit the data correlation crucial for accurate source localization. This correlation can be effectively captured using graphs, which consider the spatial relationship among data points through edges. This work introduces a novel graph learning module for source localization using spectrograms from ships-of-opportunity, which represent mid-frequency acoustic broadband signals from ship-radiated noise ranging from 360 to 1100 Hz, collected during the 2017 Seabed Characterization Experiment (SBCEX 2017). The proposed approach follows a two-step process: first, a pre-trained convolutional neural network (CNN) module is used for feature extraction via self-supervised learning, and then a graph neural network model is trained using semi-supervised learning for source localization. The graph is constructed using a k-nearest neighbor algorithm, incorporating features extracted by the CNN from the spectrograms. By employing this two-stage training strategy, our framework addresses the challenge of limited labeled data availability while achieving performance comparable to conventional supervised learning models. The effectiveness of our approach is demonstrated through model evaluation on both synthetic and measured data, showcasing the architecture's ability to generalize well to unseen scenarios. (10.1121/10.0039042)
  DOI : 10.1121/10.0039042
• Proceedings of the 5th Conference on Language, Data, and Knowledge
  
  • Alam Mehwish
  • Tchechmedjiev Andon
  • Gracia Jorge
  • Gromann Dagmar
  • di Buono Maria Pia
  • Monti Johanna
  • Ionov Maxim
  , 2025. Questo volume contiene gli atti della quinta conferenza Language Data and Knowledge (LDK), che si è tenuta a Napoli, Italia, dal 9 all’11 settembre 2025. L’evento si è svolto in modalità ibrida, con la maggior parte dei partecipanti presenti in loco. La conferenza biennale, inaugurata nel 2017, riunisce esperti in tecnologie del linguaggio, scienza dei dati e rappresentazione della conoscenza. Supportata da un comitato scientifico internazionale, LDK è cresciuta costantemente, con le edizioni precedenti ospitate in Irlanda, Germania, Spagna e Austria. Gli atti di questa edizione di LDK raccolgono 34 articoli e una prefazione. Ciascun articolo è stato sottoposto a revisione single-blind da almeno tre esperti. La conferenza si concentra sull’acquisizione e sull’uso dei dati linguistici in contesti scientifici e industriali, con particolare attenzione all’elaborazione del linguaggio naturale, all’apprendimento automatico e alle tecnologie semantiche. I temi principali includono grafi della conoscenza, risorse multilingue e approcci neuro-simbolici che combinano modelli linguistici di grandi dimensioni e semantica esplicita. (10.6093/978-88-6719-333-2)
  DOI : 10.6093/978-88-6719-333-2
• QINCODEC: Neural Audio Compression with Implicit Neural Codebooks
  
  • Lahrichi Zineb
  • Hadjeres Gaëtan
  • Richard Gael
  • Peeters Geoffroy
  , 2026. Neural audio codecs, neural networks which compress a waveform into discrete tokens, play a crucial role in the recent development of audio generative models. State-of-the-art codecs rely on the end-to-end training of an autoencoder and a quantization bottleneck. However, this approach restricts the choice of the quantization methods as it requires to define how gradients propagate through the quantizer and how to update the quantization parameters online. In this work, we revisit the common practice of joint training and propose to quantize the latent representations of a pre-trained autoencoder offline, followed by an optional finetuning of the decoder to mitigate degradation from quantization. This strategy allows to consider any off-the-shelf quantizer, especially state-of-the-art trainable quantizers with implicit neural codebooks such as QINCO2. We demonstrate that with the latter, our proposed codec termed QINCODEC, is competitive with baseline codecs while being notably simpler to train. Finally, our approach provides a general framework that amortizes the cost of autoencoder pretraining, and enables more flexible codec design.
• Bidding efficiently in Simultaneous Ascending Auctions with incomplete information using Monte Carlo Tree Search and determinization
  
  • Pacaud Alexandre
  • Bechler Aurelien
  • Coupechoux Marceau
  IEEE Transactions on Games, Institute of Electrical and Electronics Engineers, 2025, 17 (3), pp.813-826. In this paper, we tackle the problem of designing an efficient bidding strategy for Simultaneous Ascending Auctions (SAA). SAA is well-known mechanism for allocating spectrum to mobile networks operators and has been used for example to allocate 5G licenses in many countries. Although the rules are relatively simple, there is no known optimal bidding strategy for SAA. In a previous work, we proposed a Simultaneous Move Monte-Carlo Tree Search (SM-MCTS) based algorithm named SMSα that we extend here to an incomplete information framework. We consider and compare three determinization approaches of SMSα, and show how they are able to tackle four key strategic issues of SAA, namely the exposure problem, the own price effect, the budget constraints and the eligibility management. Extensive numerical experiments on instances of realistic size and including an uncertain framework show that our extensions of SMSα outperform state-of-the-art algorithms by achieving higher expected utility while taking less risks. (10.1109/TG.2025.3552025)
  DOI : 10.1109/TG.2025.3552025
• On the Effect of Feature Reduction on Energy Consumption: An Exploratory Study
  
  • Tërnava Xhevahire
  • Lefeuvre Romain
  • Perez Quentin
  • Khelladi Djamel Eddine
  • Acher Mathieu
  • Combemale Benoît
  , 2025, pp.1-11. Energy consumption is a growing concern for sustainable software. Although increasingly studied, it remains largely unexplored in configurable systems growing in complexity with features. Feature reduction can eliminate software bloat, but to our knowledge, its impact on energy use has not been investigated. To fill this gap, we investigated how both on-demand and built-in feature reduction (defined later) affect the energy consumption of configurable systems. We conducted a first exploratory study using 28 programs from three systems with built-in feature reduction, namely ToyBox, BusyBox, and GNU, as well as 6 GNU programs debloated on-demand using the Chisel, Debop, and Cov tools. In our results, built-in feature reduction led to statistically significant energy decreases in 7% of the cases, while on-demand reduction, despite achieving energy decreases in 67% of cases, showed no statistical significance. However, when energy consumption increased, it was often more substantial than the reductions observed (occurring in 25% of built-in cases and 11% of on-demand cases) showing the complex and sometimes counterintuitive interplay between feature reduction and energy. Additionally, the observed strong correlation between energy consumption and execution time motivates a shift from traditional debloating goals, centered on binary size/attack surface, to energy-aware strategies that prioritize performance concerns. Finally, we provide an in-depth analysis and discuss the perspective. (10.1145/3744915.3748463)
  DOI : 10.1145/3744915.3748463
• Quantum Reupload Units: A Scalable and Expressive Approach for Time Series Learning
  
  • Cassé Léa
  • Ponnambalam Sabarikirishwaran
  • Pfahringer Bernhard
  • Bifet Albert
  , 2025, pp.1815-1825. <div><p>We propose a single-qubit Quantum Machine Learning (QML) model for time series forecasting, built around the concept of a Quantum Reupload Unit (QRU), a hardwareefficient quantum circuit architecture with shallow depth. The proposed model demonstrates enhanced predictive power compared to variational methods such as quantum circuits (VQC), parameterized quantum circuits (PQC), and quantum residual blocks (QRB). The proposed QRU outperforms classical learning models such as Recurrent Neural Networks (RNNs) and Long-Short Term Memory (LSTM) with the same number of parameters. The novelty of this approach is its ability to model temporal patterns without relying on an extensive memory state, which reduces resource demands while preserving forecast accuracy. The expressivity of the model is evaluated through Fourier spectral decomposition. We analyze the trainability of our model using the absorption witness metric. We benchmarked the proposed model on the Mackey-Glass chaotic time series and the real-world river level dataset from TAIAO. The proposed model consistently exhibits enhanced expressivity over both of the datasets. These results highlight the significance of QRUs as promising candidates for learning models that can be conveniently deployed on noisy intermediate-scale quantum (NISQ) hardware.</p></div> (10.1109/QCE65121.2025.00199)
  DOI : 10.1109/QCE65121.2025.00199
• Federated reinforcement learning for scheduling-offloading policies in multi-cluster NOMA systems
  
  • Djemai Ibrahim
  • Sarkiss Mireille
  • Ciblat Philippe
  EURASIP Journal on Advances in Signal Processing, SpringerOpen, 2025, 2025 (37), pp.1-31. Intelligent scheduling and resource allocation of User Equipments (UEs) in wireless networks has been an ongoing topic of research. The innovation in this field focuses mostly on generalizing the system to include more components, as well as deriving new ways to solve the problem. We address in this paper an unexplored case of the scheduling-offloading problem for a wireless network with Mobile Edge Computing (MEC). In this network, the UEs have mobility models and are transmitting using Non-Orthogonal Multiple Access (NOMA). They are also equipped with data buffers and batteries with Energy Harvesting (EH) capabilities. We propose a novel UEs clustering approach to account for the growing NOMA inter-user interference, which can lead to performance issues especially in the downlink decoding phase. In addition, clustering can help reduce the problem complexity by distributing it among the clusters that operate independently. We investigate Deep Reinforcement Learning (DRL) to devise efficient policies that minimize the packet loss due to delay infringements. Moreover, we use Federated Learning (FL) to learn a unified policy accounting for the dynamic nature of clusters. Our simulation results based on DRL method, namely the Proximal Policy Optimization (PPO), and standard methods, show the effectiveness of using learning-based algorithms in terms of minimizing the packet loss and the energy consumption. (10.1186/s13634-025-01242-7)
  DOI : 10.1186/s13634-025-01242-7
• Graph Neural Networks for Moving Objects Detection in Videos
  
  • Prummel Wieke
  • Giraldo Jhony
  • Zakharova Anastasia
  • Bouwmans Thierry
  , 2025, 09, pp.121-143. Deep learning has been widely applied for the detection of moving objects from static cameras. Recently, many methods using graph neural networks for background subtraction have been reported with very promising performance. This chapter provides a survey of different graph neural for moving object detection. First, a comparison of the transductive and inductive architectures of each method is provided, followed by a discussion of the specific application requirements, such as spatio-temporal and real-time constraints. After analyzing the strategies of each method and showing their limitations, a comparative evaluation of the large-scale CDnet2014 dataset is provided. Finally, we conclude with some potential future research directions. (10.1142/9789819807154_0006)
  DOI : 10.1142/9789819807154_0006
• Blind Polarisation Demultiplexing and Carrier Recovery Using FIR-based Variational AutoEncoder Equaliser for Probabilistic Constellation Shaping in Optical Fibre Communications
  
  • Tomczyk Louis
  • Awwad Élie
  • Ware Cédric
  Journal of Lightwave Technology, Institute of Electrical and Electronics Engineers (IEEE)/Optical Society of America(OSA), 2025, pp.1-15. <div><p>We investigate through simulations the potential of Finite Impulse Response (FIR)-based Variational AutoEncoderinspired (VAE-FIR) equaliser for polarisation demultiplexing, Carrier Phase Recovery (CPR), and Carrier Frequency Offset (CFO) estimation in the context of Probabilistic Constellation Shaped (PCS) transmissions in coherent optical fibre communication systems. Additionally, we compare the performance of this novel estimator with the conventional Constant Modulus Algorithm (CMA) and Pilot-Aided Carrier Phase Recovery (PA-CPR). Our study shows that the VAE-FIR clearly outperforms the conventional approach in terms of polarisation demultiplexing, even with PCS where the CMA fails. We also show the ability of the VAE-FIR to track the phase evolution. Its ability to compensate for the carrier's phase effects is however limited to linewidths of a few dozen kHz and a hundred kHz for CFO, showing that the VAE-FIR may be used to compensate for the small residual phase noise or residual frequency mismatch.</p></div> (10.1109/JLT.2025.3603685)
  DOI : 10.1109/JLT.2025.3603685
• Super-résolution non supervisée d'images hyperspectrales de télédétection utilisant un entraînement entièrement synthétique
  
  • Xu Xinxin
  • Gousseau Yann
  • Kervazo Christophe
  • Ladjal Saïd
  , 2025. Hyperspectral single-image super-resolution (SISR) aims to improve the spatial resolution of images while preserving their spectral richness. Most current methods rely on supervised learning that requires high-resolution reference data, which are often unavailable in practice. To overcome this limitation, we propose an unsupervised learning approach based on the generation of synthetic data. The hyperspectral image is first decomposed into materials and abundances using a hyperspectral unmixing algorithm. A neural network is then trained to super-resolve the abundance maps from synthetic data generated through a dead leaves model, which imitates the statistical properties of real abundances. The high-resolution hyperspectral image is finally reconstructed by recombining the super-resolved abundance maps with the materials. Experimental results validate the effectiveness of this approach and highlight the usefulness of synthetic data for training.
• Désentrelacement fréquentiel doux pour les codecs audio neuronaux
  
  • Giniès Benoît
  • Bie Xiaoyu
  • Fercoq Olivier
  • Richard Gaël
  , 2025. Bien que les modèles basés sur les réseaux de neurones aient permis des avancées significatives dans l'extraction de représentations audio, l'interprétabilité des représentations apprises reste un défi majeur. Pour y remédier, des techniques de désentrelacement ont été intégrées dans les codecs audio neuronaux discrets afin d'imposer une structure aux tokens extraits. Cependant, ces approches sont souvent fortement dépendantes de tâches ou d'ensembles de données spécifiques. Dans ce travail, nous proposons un codec audio neuronal désentrelacé qui tire parti de la décomposition spectrale des signaux temporels pour améliorer l'interprétabilité de la représentation. Des évaluations expérimentales démontrent que notre méthode surpasse un modèle de référence en termes de fidélité de reconstruction et de qualité perceptuelle.
• Vers une axiomatique minimale de l'information quantique : Schrödinger ex-nihilo
  
  • Rioul Olivier
  , 2025. Partant de deux simples postulats quantiques, on peut retrouver l’équation de Schrödinger qui modélise l’évolution d’un système quelconque en dimension finie. Cette équation de la physique fondamentale, qui peut sembler absconse pour la communauté du traitement du signal, n’est en effet qu’une conséquence de propriétés d’orthogonalité dans un espace de Hilbert et d’une hypothèse probabiliste sur la mesure quantique. Sa preuve mathématique utilise des versions simplifiées des théorèmes de Wigner et de Stone.
• Reconstruction 3D depuis des couples SAR ascendant/descendant
  
  • Barbier--Renard Emile
  • Tupin Florence
  • Denis Loïc
  , 2025, pp.1-4. Afin d'exploiter les disparités géométriques entre des images SAR acquises sous différents points de vue, nous présentons une méthode de reconstruction de surface par rendu inverse inspirée de NeRF. Elle optimise une carte d'élévation et une carte de coefficients de rétrodiffusion à partir d'un minimum de deux images, et s'appuie sur un modèle de rendu différentiable adapté à cette représentation en carte d'élévation ainsi qu'une stratégie multi-échelles assurant une convergence rapide. Nous validons les capacités de reconstruction sur des données synthétiques réalistes générées par le simulateur EMPRISE ® de l'ONERA.
• Ordonnancement et ACM conjoint sur canal aléatoire basé sur un transformer entrainé par apprentissage profond par renforcement
  
  • Nérondat Sylvain
  • Leturc Xavier
  • Le Martret Christophe
  • Ciblat Philippe
  , 2025. Nous proposons une solution conjointe d'ordonnancement et de sélection du schéma de modulation et codage pour un système de communications sans fil sur un canal aléatoire. Les trafics de chaque lien sont "contraints en délai" ou "best effort". Les paquets sont stockés dans des files d'attente de taille finie, et les paquets arrivant alors que la file est pleine sont supprimés. Les pertes de paquets peuvent être aussi dues au canal ou à un dépassement de délai. L'allocation des ressources se fait par bloc sur une trame, chaque bloc étant dédié à un type de trafic pour un lien donné. La solution repose sur un réseau de neurones profond de type "transformer", combiné à une ramification d'action et entraîné par apprentissage par renforcement afin de minimiser la perte de paquets.
• La recherche en TdSI aux temps du transhumanisme
  
  • Maitre Henri
  , 2025, pp.1-4. <div><p>À travers des start-ups de la Tech et certains GAFAM, l'agenda transhumaniste a fait irruption dans les laboratoires de TdSI et d'IA, avec en étendard ses thèmes favoris : l'homme augmenté, la communication cerveaumachine, l'assistance mentale… Pour certains, cette confiance hardie dans le rôle de la Science pour sauver l'humanité est le meilleur défenseur de la Raison dans un monde qui doute de ses scientifiques. On montre ici le double visage des projets transhumanistes : d'une part un engagement très noble dans la résolution de quelques grands projets de la société, d'autre part une idéologie mortifère pour l'humanité qui l'engagera violemment dans une quête technosolutionniste au service d'une minorité.</p><p>Abstract -With the start-ups of Tech and GAFAM, the transhumanist agenda has burst into the laboratories of Signal &amp; Image Processing and AI, brandishing its favorite themes: augmented man, brain-machine communication, mental assistance ... For many, this bold confidence in the power of Science is the best guarantee of maintaining Reason in a world that doubts its scientists. Here we show the double face of transhumanistic projects: on the one hand a very noble commitment in the resolution of some great projects of society, on the other hand a mortifying ideology for mankind that will engage it violently in a techno-solutionist quest at the service of a minority.</p></div>
• Reconstruction 3D en tomographie radar : apprentissage profond basé sur un Matching Pursuit déroulé
  
  • Ulondu Mendes Cristiano
  • Denis Loïc
  • Kervazo Christophe
  • Tupin Florence
  , 2025. La tomographie radar en milieu urbain consiste à séparer des réflecteurs situés à des hauteurs différentes mais vus dans un même pixel car situés à une distance similaire du radar. Les méthodes d'apprentissage profond proposées récemment pour résoudre cette tâche sont basées sur le déroulement d'algorithmes de poursuites de base avec contrainte de parcimonie. Ils dépendent d'une discrétisation des hauteurs et ne permettent pas un contrôle simple du nombre de réflecteurs détectés. On présente dans cet article une approche alternative permettant d'estimer la position des cibles sur un intervalle continu. Notre approche s'inspire des itérations des algorithmes gloutons de reconstruction parcimonieuse tels que Matching Pursuit ou RELAX. Nous montrons des résultats de reconstruction satisfaisants sur des données simulées et sur une pile d'images satellitaires.
• Generalizability and sample complexity of quadratic shallow neural networks under low-rank learning
  
  • Wang Xiaolin
  • Rioul Olivier
  • Mokraoui Anissa
  • Duhamel Pierre
  • Benesty Jacob
  , 2025. We investigate the interplay between sample complexity and model complexity in low-rank learning of quadratic shallow neural networks (QSNN), within a novel doubly-correlated teacher-student framework that incorporates parameter correlations to reflect real-world data properties. This framework generalizes existing theories for QSNN by analyzing the impact of sample size on generalization loss for models under low-rank learning or exhibiting inherent bias. We observe a two-regime behavior in the scaling law of generalization ability with respect to sample size and show that parameter correlations in the teacher model significantly enhance the generalization of rank-reduced models. Extensive numerical simulations confirm the results and offer theoretical insights and practical guidance for designing efficient neural network architectures under low-rank learning.
• Une distance de style stochastique pour la synthèse de textures multispectrales
  
  • Ollivier Sélim
  • Gousseau Yann
  • Lefebvre Sidonie
  , 2025, pp.985-988/2025-1702. In this paper, we propose a novel multispectral style distance that relies on a RGB network. It consists in the stochastic evaluation of a classical style distance, over images formed by random triplets of spectral bandes. It constitutes a simple and efficient extension of state-of-the-art methods for multispectral texture synthesis, while avoiding the additional training of a multispectral network.
• Détection non supervisée de changements radiométriques en imagerie radar à synthèse d'ouverture
  
  • Bultingaire Thomas
  • Kervazo Christophe
  • Denis Loïc
  • Tupin Florence
  , 2025, pp.1-4. L'imagerie radar à synthèse d'ouverture est un mode d'imagerie clé pour la détection de changements en télédétection. Cette tâche est difficile à cause du phénomène de chatoiement, un phénomène qui nécessite de réaliser une étape de débruitage pour y être davantage robuste. Cependant, il est nécessaire de prendre en compte les incertitudes de débruitage pour contrôler la probabilité de fausse alarme des changements détectés car les instabilités de débruitage doivent être distinguées des changements. Nous proposons donc un réseau, entraîné de manière auto-supervisée, pour prédire les incertitudes de débruitage menant à une détection de changements radiométriques dont la performance est évaluée sur des images du satellite TerraSAR-X.