Publications

2025

• 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.
• 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.
• 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.
• 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.
• 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.
• SigN: SIMBox Activity Detection Through Latency Anomalies at the Cellular Edge
  
  • Kouam Anne Josiane
  • Carneiro Viana Aline
  • Martins Philippe
  • Adjih Cédric
  • Tchana Alain
  , 2025. Despite their widespread adoption, cellular networks face growing vulnerabilities due to their inherent complexity and the integration of advanced technologies. One of the major threats in this landscape is Voice over IP (VoIP) to GSM gateways, known as SIMBox devices. These devices use multiple SIM cards to route VoIP traffic through cellular networks, enabling international bypass fraud with losses of up to $3.11 billion annually. Beyond financial impact, SIMBox activity degrades network performance, threatens national security, and facilitates eavesdropping on communications. Existing detection methods for SIMBox activity are hindered by evolving fraud techniques and implementation complexities, limiting their practical adoption in operator networks. This paper addresses the limitations of current detection methods by introducing SigN , a novel approach to identifying SIMBox activity at the cellular edge. The proposed method focuses on detecting remote SIM card association, a technique used by SIMBox appliances to mimic human mobility patterns. The method detects latency anomalies between SIMBox and standard devices by analyzing cellular signaling during network attachment. Extensive indoor and outdoor experiments demonstrate that SIMBox devices generate significantly higher attachment latencies, particularly during the authentication phase, where latency is up to 23 times greater than that of standard devices. We attribute part of this overhead to immutable factors such as LTE authentication standards and Internet-based communication protocols. Therefore, our approach offers a robust, scalable, and practical solution to mitigate SIMBox activity risks at the network edge. (10.1145/3708821.3733902)
  DOI : 10.1145/3708821.3733902
• Modèle physique variationnel pour l’estimation de réponses impulsionnelles de salles
  
  • Lalay Louis
  • Fontaine Mathieu
  • Badeau Roland
  , 2025. Estimer la réponse impulsionnelle d’une salle est essentiel pour des tâches comme la déréverbération, qui améliore la reconnaissance automatique de la parole. La plupart des méthodes existantes reposent soit sur du traitement du signal statistique, soit sur des réseaux de neurones profonds s’inspirant du traitement du signal. Cependant, la combinaison des modélisations statistique et physique reste largement inexploré en estimation de réponse impulsionnelle de salle. Cet article propose une approche novatrice intégrant les deux aspects à travers un modèle physique. La réponse de salle est décomposée en paramètres interprétables : un bruit blanc gaussien modulé par une décroissance exponentielle dépendante de la fréquence (modélisant l’absorption des murs) et un filtre autorégressif (modélisant par exemple la réponse du microphone). L’optimisation d’une fonction d’énergie libre variationnelle permet une estimation pratique des paramètres. Nous montrons que, connaissant les signaux secs et réverbérants, la méthode proposée surpasse la déconvolution classique dans des environnements bruités, comme le confirment les mesures objectives.
• Déréverbération non-supervisée de la parole par modèle hybride
  
  • Bahrman Louis
  • Fontaine Mathieu
  • Richard Gaël
  , 2025, pp.1-4. Cet article introduit une nouvelle stratégie d'apprentissage pour améliorer des systèmes de déréverbération de la parole de manière non-supervisée en n'utilisant que des signaux réverbérants. La plupart des algorithmes existants nécessitent des paires de signaux (sec, réverbérant), qui sont difficiles à obtenir. Notre approche utilise en revanche des informations acoustiques limitées, comme le temps de réverbération (RT60), pour entraîner un système de déréverbération. Les résultats expérimentaux démontrent que notre méthode permet d'obtenir des performances plus cohérentes que l'état de l'art sur différentes mesures objectives.
• La borne bayesienne de Schützenberger-van Trees : Un principe d'incertitude sur l'a posteriori
  
  • Rioul Olivier
  • Renaux Alexandre
  , 2025. Cette communication propose une perspective historique sur la borne de Cramér-Rao bayésienne (BCRB), généralement attribuée à van Trees qui l’a découverte en 1968. Selon la loi de Stigler sur l’éponymie, aucune découverte scientifique ne porte le nom de son premier découvreur. C’est non seulement le cas de la borne Cramér-Rao elle-même – due notamment aux mathématiciens français Fréchet et Darmois – mais aussi de l’inégalité de van Trees. En effet, le médecin, généticien, épidémiologiste et mathématicien français Marcel-Paul (Marco) Schützenberger, dans un petit article d’une quinzaine de lignes seulement, écrit en 1956 – plus d’une décennie avant van Trees – avait non seulement démontré la BCRB, mais comme le montre une lecture approfondie de sa preuve, l’avait fait avec une démarche très originale en la reliant au principe d’incertitude de Weyl-Heisenberg sur l’a posteriori. Nous passons en revue et comparons les contributions de Schützenberger et de van Trees ainsi que celles de Gart en 1959. L’équivalence générale entre BCRB et principe d’incertitude sur l’a posteriori ouvre de nouvelles perspectives.
• 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.
• Gaussian Process Regression of Steering Vectors With Physics-Aware Deep Composite Kernels for Augmented Listening
  
  • Di Carlo Diego
  • Koyama Shoichi
  • Aditya Arie Nugraha
  • Mathieu Fontaine
  • Yoshiaki Bando
  • Kazuyoshi Yoshii
  , 2025. This paper investigates continuous representations of steering vectors over frequency and position of microphone and source for augmented listening (e.g., spatial filtering and binaural rendering) with precise control of the sound field perceived by the user. Steering vectors have typically been used for representing the spatial characteristics of the sound field as a function of the listening position. The basic algebraic representation of steering vectors assuming an idealized environment cannot deal with the scattering effect of the sound field. One may thus collect a discrete set of real steering vectors measured in dedicated facilities and super-resolve (i.e., upsample) them. Recently, physics-aware deep learning methods have been effectively used for this purpose. Such deterministic super-resolution, however, suffers from the overfitting problem due to the non-uniform uncertainty over the measurement space. To solve this problem, we integrate an expressive representation based on the neural field (NF) into the principled probabilistic framework based on the Gaussian process (GP). Specifically, we propose a physics-aware composite kernel that model the directional incoming waves and the subsequent scattering effect. Our comprehensive comparative experiment showed the effectiveness of the proposed method under data insufficiency conditions. In downstream tasks such as speech enhancement and binaural rendering using the simulated data of the SPEAR challenge, the oracle performances were attained with less than ten times fewer measurements.
• Unified Variational and Physics-aware Model for Room Impulse Response Estimation
  
  • Lalay Louis
  • Fontaine Mathieu
  • Badeau Roland
  , 2025. Room impulse response estimation is essential for tasks like speech dereverberation, which improves automatic speech recognition. Most existing methods rely on either statistical signal processing or deep neural networks designed to replicate signal processing principles. However, combining statistical and physical modeling for RIR estimation remains largely unexplored. This paper\footnote{This paper was submitted to Interspeech 2025} proposes a novel approach integrating both aspects through a theoretically grounded model. The RIR is decomposed into interpretable parameters: white Gaussian noise filtered by a frequency-dependent exponential decay (e.g. modeling wall absorption) and an autoregressive filter (e.g. modeling microphone response). A variational free-energy cost function enables practical parameter estimation. As a proof of concept, we show that given dry and reverberant speech signals, the proposed method outperforms classical deconvolution in noisy environments, as validated by objective metrics.
• Hardness of M-LWE with General Distributions and Applications to Leaky Variants
  
  • Boudgoust Katharina
  • Jeudy Corentin
  • Tairi Erkan
  • Wen Weiqiang
  , 2025. The Module Learning With Errors (M-LWE) problem has become a fundamental hardness assumption for lattice-based cryptography. It offers an attractive trade-o between strong robustness guarantees, sometimes directly based on worst-case lattice problems, and efficiency of the subsequent cryptographic primitives. Different flavors of M-LWE have then been introduced towards improving performance. Such variants look at different secret-error distributions and might allow for additional hints on the secret-error vector. Existing hardness results however only cover restricted classes of said distributions, or are tailored to specific leakage models. This lack of generality hinders the design of efficient and versatile cryptographic schemes, as each new distribution or leakage model requires a separate and nontrivial hardness evaluation In this work, we address this limitation by establishing the hardness of M-LWE under general distributions. As a first step, we show that M-LWE remains hard when the error vector follows an arbitrary bounded distribution with sufficient entropy, with some restriction on the number of samples. Building on this, we then reduce to the Hermite Normal Form (HNF) where the secret-error vector follows said arbitrary distribution. Overall, our result shows the actual shape of the distribution does not matter, as long as it keeps sufficient entropy. To demonstrate the versatility of our framework, we further analyze a range of leakage scenarios. By examining the residual entropy given the leakage, we show that our results of M-LWE with general distributions encompass various types of leakage. More precisely, we cover exact and approximate linear hints which are widely used in recent cryptographic designs, as well as quadratic, and even non-algebraic forms, some of which were not yet covered by any theoretical hardness guarantees. The generality of our results aims at facilitating future cryptographic designs and security analyses.
• Emotional speech markers of psychiatric disturbance in Huntington’s disease
  
  • Chenain Lucie
  • Fabre Audrey
  • Titeux Hadrien
  • Morgado Graça
  • Youssov Katia
  • Clavel Chloé
  • Bachoud-Lévi Anne-Catherine
  Frontiers in Psychiatry, Frontiers, 2025, 16, pp.1633492-1:1633492-20. Introduction: Psychiatric disorders and difficulties in emotional expression represent a major problem in the management of Huntington's Disease (HD). To improve patient follow-up, we propose to investigate the link between emotional expression and psychiatric symptoms, measured by the Problem Behaviors Assessment (PBA) scale. To this aim we developed the first emotional/psychiatric speech corpus, emoHD. Methods: We included 102 HD gene carriers and 35 healthy controls (HC). Psychiatric symptoms were assessed using PBA sub-scales for Depression, Irritability/aggressivity, Apathy, and Obsessive/compulsive symptoms. Speech was annotated using three emotional descriptors: primary emotions, affective phenomena, and activation levels. Affective phenomena labels were selected based on PBA statements by external participants unaware of the study's aims. We analyzed (1) emotional descriptors' relationships, (2) emotional expression differences between HD and HC, and (3) the associations between emotions and psychiatric symptoms. Results: HD patients showed reduced emotional expressiveness than HC with more neutral activation levels (=0). Only the primary emotion "angry" was less expressed in HD compared to HC. In contrast they expressed more affective phenomena states like apathetic, confused, "depressed", "disoriented", "frustrated", and "pessimistic" than HC, whereas they expressed less "other" and "irritable" than HC. Expressed emotions were congruent with psychiatric symptoms (e.g., "anxious" and "nervous" are positively associated with Depression PBA sub-scale; "frustrated" with Irritability/aggressivity PBA sub-scale). (10.3389/fpsyt.2025.1633492)
  DOI : 10.3389/fpsyt.2025.1633492
• Efficient Negative Weight Realization for Analog Resistive Neural Networks
  
  • Kiraz Zulal
  • Pham Dang-Kièn Germain
  • Desgreys Patricia
  , 2025. Most analog nonlinear resistive neural networks for machine learning training use doubling input and output neuron nodes to implement negative weights. However, this approach increases network size, modifies the gradient computation, and complicates circuit design. We propose an alternative circuit topology that retains a one-to-one correspondence between neurons in the original model and their analog counterparts. Our design employs a emph{single} input source for all first-layer weights, a emph{single} resistor per weight, and a bidirectional amplifier for the rest of the layers' weight to handle negative connections without duplicating neurons. We validate our design on a binary XOR classification task over SI{100}{} training epochs and SI{100}{} randomized initializations. Our textbf{single-resistor} approach achieved an average final error of SI{-6.6}{dB} and required approximately SI{568}{} minutes of total CPU time. In comparison, the textbf{doubled-node} design reached SI{-4.6}{dB} error and consumed around SI{1104}{} minutes of CPU time. This equates to nearly 49% less computation for the single-resistor circuit while preserving the standard gradient update procedure—demonstrating that negative weights can be realized more efficiently without doubling input/output neurons.
• Immersed boundary–lattice Boltzmann mesoscale method for wetting problems
  
  • Bellantoni Elisa
  • Guglietta Fabio
  • Pelusi Francesca
  • Desbrun Mathieu
  • Um Kiwon
  • Nicolaou Mihalis
  • Savva Nikos
  • Sbragaglia Mauro
  Physical Review E, American Physical Society (APS), 2025, 112 (2), pp.025305 (1-15). We develop a mesoscale computational model to describe the interaction of a droplet with a solid. The model is based on the hybrid combination of the immersed boundary and the lattice Boltzmann computational schemes: the former is used to model the non-ideal sharp interface of the droplet coupled with the inner and outer fluids, simulated with the lattice Boltzmann scheme. We further introduce an interaction force to model the wetting interactions of the droplet with the solid at mesoscale: this interaction force is designed with the key computational advantage of providing a regularization of the interface profile close to the contact line, avoiding abrupt curvature changes that could otherwise cause numerical instabilities. The proposed model substantially improves earlier immersed boundary - lattice Boltzmann models for wetting in that it allows a description of an ample variety of wetting interactions, ranging from hydrophobic to hydrophilic cases, without the need for any pre-calibration study on model parameters to be used. Model validations against theoretical results for droplet shape at equilibrium and scaling laws for droplet spreading dynamics are addressed. (10.1103/mp3p-8j22)
  DOI : 10.1103/mp3p-8j22
• Time-resolved second-order autocorrelation function of parametric down-conversion
  
  • Horoshko Dmitri
  • Srivastava Shivang
  • Sośnicki Filip
  • Mikołajczyk Michał
  • Karpiński Michał
  • Brecht Benjamin
  • Kolobov Mikhail
  Physical Review A, American Physical Society, 2025, 112 (2), pp.023703-1:023703-13. We study a possibility of measuring the time-resolved second-order autocorrelation function of one of two beams generated in type-II parametric down-conversion by means of temporal magnification of this beam, bringing its correlation time from the picosecond to the nanosecond scale, which can be resolved by modern photodetectors. We show that such a measurement enables one to infer directly the degree of global coherence of that beam, which is linked by a simple relation to the number of modes characterizing the entanglement between the two generated beams. We illustrate the proposed method by an example of photon pairs generated in a periodically poled potassium titanyl phosphate (KTP) crystal with a symmetric group velocity matching for various durations of the pump pulse, resulting in different numbers of modes. Our theoretical model also shows that the magnified double-heralded autocorrelation function of one beam exhibits a local maximum around zero delay time, corresponding to photon bunching at a short time scale. (10.1103/7ckm-tm3r)
  DOI : 10.1103/7ckm-tm3r
• SpectreShield: Design and Analysis of Spectre Countermeasures on RISC-V Using gem5
  
  • Khan Mahreen
  • Mushtaq Maria
  • Pacalet Renaud
  • Apvrille Ludovic
  , 2025. <div><p>Speculative execution attacks like Spectre exploit microarchitectural side effects to leak sensitive data during transient execution. While various software and hardware countermeasures have been proposed for x86 and ARM architectures, their effectiveness and microarchitectural impact remain underexplored on RISC-V platforms. To study such attacks and evaluate these countermeasures, simulation tools like the gem5 simulator provide detailed insights into microarchitectural state changes during speculation. In this paper, we present the first comprehensive evaluation of Spectre-v1 countermeasures on the RISC-V architecture using the gem5 full-system simulator. We implement and assess four Spectre-v1 mitigations: index masking (CM1), randomized offset (CM2), fence-based serialization (CM3), and bitwise selection (CM4). Our experiments reveal that, in the absence of mitigations, Spectre-v1 enables 100% secret key recovery. In contrast, all proposed countermeasures reduce the recovery rate to below 1%, with branch mispredictions decreasing by 41.7%-46.3%. The paper analyzes the securityperformance trade-offs of each approach. Beyond demonstrating their effectiveness, we quantify their microarchitectural impact, measuring reductions in squashed instructions, DRAM latency variability, and return address stack mispredictions. This paper provides a practical framework for evaluating transient execution defenses and advances secure-by-design RISC-V processors.</p></div>
• Bayesian Stream Tuner: Dynamic Hyperparameter Optimization for Real-Time Data Streams
  
  • Verma Nilesh
  • Bifet Albert
  • Pfahringer Bernhard
  • Bahri Maroua
  , 2025, 2, pp.2871-2882. Hyperparameter optimization is crucial for maximizing machine learning model performance, yet most existing algorithms are designed for batch or offline scenarios and assume static data distributions. Such assumptions fall short in data stream settings, where models must adapt to evolving inputs in real time. To address these limitations, we propose the Bayesian Stream Tuner (BST), a novel framework for online hyperparameter optimization in nonstationary data streams. BST maintains a dynamic set of candidate hyperparameter configurations and periodically refines them using an incremental Bayesian model, which estimates configuration performance based on recent data statistics and hyperparameter values. This systematic exploration and refinement strategy allows BST to detect and respond to concept drift by resetting its adaptation mechanisms whenever necessary, ensuring strong performance under changing distributions. Our theoretical analysis establishes sublinear regret bounds for BST in dynamic environments, and extensive experiments on classification and regression tasks demonstrate that BST consistently outperforms state-of-the-art online hyperparameter optimization methods in both predictive accuracy and adaptability, making it a powerful solution for real-time hyperparameter tuning in evolving data streams. (10.1145/3711896.3736852)
  DOI : 10.1145/3711896.3736852
• Satellite Image Time-Series Data Augmentation Using an Attention Mechanism Variational Recurrent Autoencoder
  
  • Chaabane Ferdaous
  • Tupin Florence
  , 2025. Data scarcity presents a significant challenge in satellite image analysis, particularly for developing robust models in remote sensing applications. High-quality and abundant data are essential for accurate predictions; however, acquiring Satellite Image Time-Series (SITS) data is often constrained by factors such as limited temporal coverage and the high cost of Very High Resolution (VHR) acquisitions. To address this issue, we propose a novel Attention-based Variational Recurrent Autoencoder (AVRAE) designed for generating synthetic satellite image time-series data. This method extends the evidence lower bound (ELBO) of variational inference to incorporate the temporal dependencies essential for satellite data. A recurrent neural network-based autoencoder framework is employed, integrated with an attention mechanism to effectively capture both short-and long-term temporal relationships. The AVRAE framework synthesizes realistic and statistically representative satellite time-series data, enabling enhanced analysis for remote sensing applications. Evaluations using real-world satellite datasets demonstrate that AVRAE produces coherent and statistically valid synthetic data, thereby improving VHR SITS data quality for deep learning-based remote sensing applications.
• An Information Theoretic Proof of the Chernoff-Hoeffding Inequality
  
  • Rioul Olivier
  • Solé Patrick
  Information Processing Letters, Elsevier, 2025, 190, pp.106582. The Chernoff bound is a well-known upper bound on the tail of binomial distributions of parameter 1/2 involving the binary entropy function. Hoeffding's inequality (or the Chernoff-Hoeffding inequality) is a generalization for binomial distributions of parameter 1 -1/q, involving the q-ary entropy function (with q ≥ 2), which can be written in terms of the Kullback-Leibler divergence and is related to the bound in Fano's inequality. We give an information theoretic proof of that bound, and sketch some applications to channel and source coding. We also derive a refined bound which is always sharper. (10.1016/j.ipl.2025.106582)
  DOI : 10.1016/j.ipl.2025.106582
• Long run convergence of discrete-time interacting particle systems of the McKean-Vlasov type
  
  • Bianchi Pascal
  • Hachem Walid
  • Priser Victor
  Stochastic Processes and their Applications, Elsevier, 2025. We consider a discrete-time system of n coupled random vectors, a.k.a. interacting particles. The dynamics involve a vanishing step size, some random centered perturbations, and a mean vector field which induces the coupling between the particles. We study the doubly asymptotic regime where both the number of iterations and the number n of particles tend to infinity, without any constraint on the relative rates of convergence of these two parameters. We establish that the empirical measure of the interpolated trajectories of the particles converges in probability, in an ergodic sense, to the set of recurrent Mc-Kean-Vlasov distributions. A first application example is the granular media equation, where the particles are shown to converge to a critical point of the Helmholtz energy. A second example is the convergence of stochastic gradient descent to the global minimizer of the risk, in a wide two-layer neural networks using random features.
• On the spectral decomposition of the complex Robin Laplacian
  
  • Badeau Roland
  Journal of the Acoustical Society of America, Acoustical Society of America, 2025, 158 (1), pp.838-848. The mathematical properties of the Laplacian on a bounded domain are well-known when the boundary condition is of the first type (Dirichlet) or second type (Neumann). In both cases, this operator is self-adjoint and, therefore, diagonalizable, its spectrum is discrete, and the set of eigenfunctions can be chosen to form an orthonormal basis of the Hilbert space of square-integrable functions on the domain. However, in the case of the third type (Robin) boundary condition, the same is true only when the parameter is real-valued. On the contrary, when this parameter is complex-valued, the Laplacian may not even be diagonalizable. In this paper, the spectral decomposition of the complex Robin Laplacian is investigated in the most general case possible, and a formula that decomposes any square-integrable function on the set of its (generalized) eigenfunctions is provided. This result is applied to the Green's function of the Helmholtz equation, whose existence, unicity, and closed-form expression are established in this general setting, and the statistical wave field theory, which provides the statistical laws of waves propagating in a bounded domain. (10.1121/10.0037233)
  DOI : 10.1121/10.0037233
• Melody-Lyrics Matching with Contrastive Alignment Loss
  
  • Wang Changhong
  • Olvera Michel
  • Richard Gaël
  , 2025. The connection between music and lyrics is far beyond semantic bonds. Conceptual pairs in the two modalities such as rhythm and rhyme, note duration and syllabic stress, and structure correspondence, raise a compelling yet seldom-explored direction in the field of music information retrieval. In this paper, we present melody-lyrics matching (MLM), a new task which retrieves potential lyrics for a given symbolic melody from text sources. Rather than generating lyrics from scratch, MLM essentially exploits the relationships between melody and lyrics. We propose a self-supervised representation learning framework with contrastive alignment loss for melody and lyrics. This has the potential to leverage the abundance of existing songs with paired melody and lyrics. No alignment annotations are required. Additionally, we introduce sylphone, a novel representation for lyrics at syllable-level activated by phoneme identity and vowel stress. We demonstrate that our method can match melody with coherent and singable lyrics with empirical results and intuitive examples. We open source code and provide matching examples on the companion webpage: https://github.com/changhongw/mlm.
• Benchmarking the Benchmarks: Reproducing Climate-Related NLP Tasks
  
  • Calamai Tom
  • Balalau Oana
  • Suchanek Fabian M
  , 2025. Significant efforts have been made in the NLP community to facilitate the automatic analysis of climate-related corpora by tasks such as climate-related topic detection, climate risk classification, question answering over climate topics, and many more. In this work, we perform a reproducibility study on 8 tasks and 29 datasets, testing 6 models. We find that many tasks rely heavily on surface-level keyword patterns rather than deeper semantic or contextual understanding. Moreover, we find that 96% of the datasets contain annotation issues, with 16.6% of the sampled wrong predictions of a zero-shot classifier being actually clear annotation mistakes, and 38.8% being ambiguous examples. These results call into question the reliability of current benchmarks to meaningfully compare models and highlight the need for improved annotation practices. We conclude by outlining actionable recommendations to enhance dataset quality and evaluation robustness.