Séminaire CORNET – Roseli Wedemann – 09/11/2026

9 novembre 2026

Un séminaire Cornet de Roseli Wedemann (UERJ, Brésil) aura lieu le lundi 09/11 à 11h30 Salle 5 CERI : ================== Titre: « Consciousness-Unconsciousness, Associative Memory and Generalized Statistical Mechanics » We will present a review of our modelling efforts, where we have described mental pathology [1, 2, 3] considering its relation to memory function, and proposed neural network mechanisms whereby neurotic behavior is described as an associative memory process in the brain [4, 5]. In our model, modules corresponding to sensorial and symbolic memories interact, representing unconscious and conscious mental processes [6, 7, 8]. We initially modelled associative memory by a Boltzmann Machine (BM) [9, 10]. The observed q-exponential and asymptotic, power-law behavior for the node-degree distributions of the self-organized network topologies generated by our model suggest that memory dynamics and associativity may not be well described by Boltzmann-Gibbs (BG) statistical mechanics [11]. We thus modeled the dynamics of memory access by a generalization of the BM called Generalized Simulated Annealing (GSA) [12], derived from the nonextensive statistical mechanical formalism [11]. Results obtained from simulations and fMRI measurements indicate that the nonextensive statistical mechanics formalism may be more appropriate to model many complex networks that constitute brain and mental structure, than the Plus d'infos

Séminaire CORNET – Hajar El Hammout – 12/10/2026

9 octobre 2026

Speaker : Professor Hajar El Hammouti from UM6P Date: 12/10/2026 – 11h30 Place: LIA meeting room. Title: “Energy Efficient Federated Learning” Abstract : Federated learning (FL) is a distributed learning framework that operates effectively over wireless networks. It enables devices to collaboratively train a model over wireless links by sharing model parameters rather than personal data. However, a key challenge in FL arises from the limited computational and communication resources of devices. As a result, energy efficiency has become a critical concern for real-world deployments. In this talk, we provide an overview of key strategies to reduce computational and communication overhead in FL systems. We will also present our recent contributions in this area, including the design of novel algorithms for client selection, model compression, and quantization. These techniques aim to minimize energy consumption while preserving model accuracy and ensuring FL convergence. Keywords: Energy Efficiency; Federated Learning; Internet of Things, Optimization; Wireless Communications. Bio:  Hajar El Hammouti is an Assistant Professor at the College of Computing of Mohammed VI Polytechnic University (UM6P) in Morocco. She received her Ph.D. in Electrical Engineering from the National Institute of Posts and Telecommunications (INPT), Rabat, and held postdoctoral positions at the International University of Rabat (UIR) and King Plus d'infos

Séminaire SLG – Maryem Bouziane – 08/10/2026

21 septembre 2026

Titre: Apprentissage de plusieurs représentations d’attributs au niveau de l’énoncé avec un encodeur de parole unifié Date: 08/10/2026 – 13h Résumé: Les modèles de fondation pour la parole entraînés par apprentissage auto-supervisé produisent des représentations génériques de la parole, capables de soutenir un large éventail de tâches de traitement de la parole. Lorsqu’ils sont ensuite adaptés par apprentissage supervisé, ces modèles peuvent atteindre de fortes performances sur des tâches aval spécifiques. Des approches récentes de post-entraînement, telles que SAMU-XLSR et SONAR, alignent les représentations de la parole sur des représentations sémantiques au niveau de l’énoncé, permettant des applications multimodales (parole–texte) et multilingues efficaces. Alors que les modèles de fondation pour la parole apprennent typiquement des embeddings contextuels au niveau de la trame acoustique, ces méthodes apprennent des représentations au niveau de l’énoncé. Dans ce travail, nous étendons ce paradigme à des attributs arbitraires au niveau de l’énoncé et proposons un cadre unifié de post-entraînement permettant à un unique modèle de fondation pour la parole de générer plusieurs types de représentations au niveau de l’énoncé. Nous démontrons l’efficacité de cette approche en apprenant conjointement des représentations sémantiques et des représentations de locuteur, ainsi que des représentations de langue et d’émotion, puis Plus d'infos

Séminaire SLG – Hugo Leguillier – 19/06/2026

16 juin 2026

Date: Friday, June 19, from 1:00 p.m. Place: room S5 Speaker: Hugo Leguillier Title: On Low-Bit Quantization Errors in Speaker Verification: Diagnostic and Mitigation Abstract: Although low-bit quantization provides practical means to deploy speaker verification on resource-constrained devices, its effects on speaker verification performance remain poorly understood. In this paper, we study uniform K-means quantization-aware training of ResNet-36 and ResNet-200 through joint layer-wise and score-level analyses. Our layer-wise analysis highlights fragile components and shows that score degradation is not fully explained by weight distortion alone. We identify a clear knee point at 2 bits, with larger score drift and harmful decision flips concentrated near the FP32 threshold. Our score-level analysis reveals where and how score errors emerge under extreme quantization. Building on these findings, we propose a calibrated multi-precision cascade that resolves most trials at 2 bits and escalates only ambiguous cases, achieving performance close to FP32 while preserving the efficiency benefits of low-bit inference with substantially lower compute and memory costs.

Séminaire SLG – Hugo Daumain – 12/06/2026

10 juin 2026

Date: Friday, June 12, from 1:00 p.m. Lieu : Salle S5 Speaker: Hugo Daumain Title: From Self-Supervised Speech Models to Mixture-of-Experts for Robust Anti-Spoofing Abstract: Recent advances in speech generation have significantly improved the naturalness of synthetic speech, making spoofing detection increasingly challenging. A key limitation of current anti-spoofing systems is their limited robustness to unseen synthesis methods. In this work, we transform a self-supervised speech representation model into a Mixture-of-Experts (MoE) architecture to improve generalization. Feed-forward blocks in selected encoder layers are replaced by multiple expert networks controlled by a layer-wise gating mechanism, allowing experts to capture complementary acoustic patterns while preserving the representations learned during self-supervised pretraining. We further analyze the architectural choices affecting the performance of this MoE conversion and investigate the activation behavior of the experts. The proposed approach is evaluated on 14 spoofing datasets and reduces the macro EER from 5.46% to 4.81%, corresponding to 11.9% relative improvement over the baseline.

Séminaire SLG – Orane Dufour – 12/06/2026

10 juin 2026

Date: Friday, June 12, from 1:00 p.m. Lieu: Salle S5 Speaker: Orane Dufour Title: A Large-Scale Per-Speaker Analysis of Re-identification Risk in Speech Anonymization Abstract: Speech anonymization is commonly evaluated using average-case metrics such as the equal error rate, which can hide large disparities in re-identification risks across individuals. In this paper, we conduct a large-scale per-speaker privacy analysis using a linkability-based metric under a worst-case scenario. Nearly 5,000 speakers are evaluated across multiple anonymization systems, attacker architectures, and conversation lengths. While linkability scores are highly polarized at the speaker level, the sets of easy to re-identify and hard to re-identify speakers vary substantially across configurations. We show that no single factor explains speaker vulnerability. Instead, the re-identification risk emerges from the interaction between the attacker, the anonymizer, and the amount of available speech. These results challenge the notion of intrinsic speaker-level privacy risks and emphasize the need for evaluation protocols that are explicitly conditioned on the attacker and anonymizer.

Séminaire CORNET – Cleque Marlain Mboulou Moutoubi – 22/05/2026

22 mai 2026

Date: 22 Mai 2026, 11h30 Lieu: Salle C057 Titre: Repeated Multi-Resource Proportional Allocation Auction Games Résumé : In the multi-resource Kelly mechanism, players obtain a share of each resource in proportion to the bids they place on it. They thus engage in a non-cooperative game where they distribute their budgets across multiple resources. In this paper, we study the repeated variant of this game under standard no-regret algorithms, namely Online Gradient Descent (OGD) and Dual Averaging (DA) algorithms. More specifically, we investigate an additive utility framework with heterogeneous valuations across resources, where each resource-specific utility can be either logarithmic or linear. In this setting, we prove uniqueness of the Nash equilibrium. Moreover, we prove convergence of OGD and DA in the repeated game to this unique Nash Equilibrium. Extensive numerical simulations validate the theoretical results and measure convergence speed across different settings.

Séminaire CORNET – Remy Kessler – 29/04/2026

13 avril 2026

TITRE: Désambiguïsation sémantique d’ontologie par LLM Date: 29 Avril 2026, 11h30 Lieu: C057 Résumé : Je présenterais dans ce séminaire les travaux réalisés l’an dernier par Anastasiia Ribova. Je présenterais une approche qui explore le potentiel des grands modèles de langage (LLM) pour automatiser la validation et l’enrichissement d’ontologies et de graphes de connaissances. À partir d’un prompt simple et en exploitant des modèles de langage de taille intermédiaire, nous avons expérimenté plusieurs stratégies, notamment l’apprentissage zero-shot, le few-shot, le raisonnement par chaîne de pensée (CoT) et le décodage par auto-cohérence (SC). Notre méthode d’ensemble a systématiquement surpassé tous les modèles individuels sur deux domaines distincts de l’ontologie. Les résultats soulignent la capacité des LLM à opérer des distinctions nuancées et à évaluer les relations dérivées de l’ontologie avec un haut niveau de précision.

Séminaire SLG – Matthew Wiesner – 26/03/2026

26 mars 2026

Salle 5 – 12h00 Titre: Modélisation Extensible de Langues et d’Accents Résumé: Les modèles d’identification de langue (LID) à l’état de l’art fonctionnent de manière fiable pour une centaine de langues. Cependant, derrière le concept de langue se cache de nombreuses variations émanant d’accents et de dialectes divers. Il est tout simplement impossible d’annoter les données en prenant compte de toutes ces variations. De plus, la parole accentuée engendre un comportement inattendu des modèles de LID et très peu de données annotées existent pour pallier le problème. Ce manque d’annotation empêche aussi l’augmentation de donnée via la synthèse de parole accentuée. Ce séminaire aborde ces problématiques et propose une ébauche de solution fondée sur une collecte de données à grande échelle à partir de diffusions radiophoniques. Cela permet d’associer aux données des annotations indirectes sous forme de géolocalisations. Le séminaire explore ensuite le lien entre la robustesse aux accents et la capacité à modéliser des séquences. Enfin, nous montrons comment ces modèles permettent d’améliorer les modèles de LID, en particulier sur la parole accentuée, et de faciliter l’extraction automatique de données accentuées pour entrainer des systèmes de synthèse vocale.Bio: Matthew Wiesner est un chercheur à Johns Hopkins University et chercheur  at Plus d'infos

Séminaire CORNET – Younes Ben Mazziane – 20/03/2026

16 mars 2026

Date: 20/03/2026 à 11h30 Lieu: Salle C057 Titre: Learning in Proportional Allocation Auctions Games Résumé : The Kelly or proportional allocation mechanism is a simple and efficient auction-based scheme that distributes an infinitely divisible resource proportionally to the agents’ bids. When agents are aware of the allocation rule, their interactions form a game, that has been extensively studied. This paper examines the less explored repeated Kelly game, focusing mainly on utilities that are logarithmic in the allocated resource fraction. We first derive this logarithmic form from fairness–throughput trade-offs in wireless network slicing, and then prove that the induced stage game admits a unique Nash equilibrium (NE). For the repeated play, we prove convergence to this NE under three behavioral models: (i) all agents use Online Gradient Descent (OGD), (ii) all agents use Dual Averaging with a quadratic regularizer (DAQ) (a variant of the Follow-the-Regularized leader algorithm), and (iii) all agents play myopic best responses (BR). Our convergence results hold even when agents use personalized learning rates in OGD and DAQ (e.g., tuned to optimize individual regret bounds), and they extend to a broader class of utilities that meet a certain sufficient condition. Finally, we complement our theoretical results with extensive Plus d'infos

1 2 3 … 7