BEGIN:VCALENDAR CALSCALE:GREGORIAN VERSION:2.0 X-WR-TIMEZONE:Europe/Paris METHOD:PUBLISH PRODID:-//LIP6//www.lip6.fr//FR X-WR-CALNAME;VALUE=TEXT:Séminaire AID X-LIC-LOCATION:Europe/Paris BEGIN:VTIMEZONE TZID:Europe/Paris BEGIN:DAYLIGHT TZOFFSETFROM:+0100 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU DTSTART:19810329T020000 TZNAME:GMT+02:00 TZOFFSETTO:+0200 END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:+0200 RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU DTSTART:19961027T030000 TZNAME:GMT+01:00 TZOFFSETTO:+0100 END:STANDARD END:VTIMEZONE BEGIN:VEVENT SUMMARY:Troisième journée des doctorants de l’axe IA&SD ORGANIZER;CN=:MAILTO: DESCRIPTION:L’objectif de cette demi-journée est de permettre aux doctorants de présenter et faire connaître leur sujet de thèse ain si que de pouvoir échanger sur leurs travaux. Ces présentations son t ouvertes à tous les membres du LIP6, permanents et non-permanents. Une visio pourra être mise en place afin de diffuser les exposés. L’après-midi se terminera par un moment convivial autour d’un g oûter. Voici les orateurs prévus et le programme pour la journé e du 5 juin (15-20mn de présentation + 10mn de questions) : - 13 h30-14h : Ioannis Kaskampas (CIAN) Title: Federated Learning on Neu romorphic Nodes Abstract: Federated learning (FL) allows multiple cl ients to train a shared model without exchanging raw data, addressing privacy and communication concerns in distributed settings. Spiking neural networks (SNNs) are well suited to energy-constrained edge dev ices because they communicate via discrete spikes rather than dense a ctivations, and dedicated neuromorphic chips like Intel Loihi can exe cute them at a fraction of the power of conventional GPUs. Most exist ing studies, focus on a single FL paradigm and a single SNN implement ation. This work studies federated learning on neuromorphic nodes acr oss several FL paradigms — synchronous, hierarchical, asynchronous, and vertical — and across heterogeneous combinations of SNN backen ds and architectures. We evaluate the resulting trade-offs in accurac y, communication cost, and stability on four event-based benchmarks. - 14h-14h30 : Martin Gomez Abejon (DECISION) Titre : Uncertainty Quantification for Pretrained Models in Time Series Résumé : L'ap prentissage profond appliqué aux séries temporelles continue d'évo luer, et un sujet de recherche particulièrement actif actuellement e st la conception, entraînement et utilisation de modèles profonds e t préentraînés sur des jeux de données de grande taille pour rés oudre des tâches sur d'autres jeux de données sans entraînement ou adaptation préalables. Pourtant, il y a toujours beaucoup de questi ons ouvertes, et certains modèles ne génèrent que des prédictions ponctuelles. Dans cette présentation, nous définissons un cadre t héorique pour la prédiction des séries temporelles, en donnant des propriétés sur les fonctions utilisées pour évaluer la quantific ation de l'incertitude, et nous étudions deux modèles de séries te mporelles récents, en comparant les intervalles qu'ils génèrent av ec ceux générés par des techniques de prédiction conforme. - 14h30-15h : Aya Tounsi (MOCAH) Titre : Optimisation multi-objectifs v isant à concilier l'état physiologique et les résultats d'apprenti ssage dans l'enseignement supérieur de longue durée Résumé : "L' apprentissage individuel reste un défi malgré l'abondance de conten us pédagogiques : comment adapter ressources et feedback pour maximi ser l'engagement, minimiser l'abandon et éviter surcharge cognitive ou ennui (état de flux, Csikszentmihalyi, 2014) ? Cette thèse intè gre exploration de données éducationnelles, psychologie et physiolo gie pour développer des algorithmes IA dans un système de gestion d e l'apprentissage. Question centrale : comment optimiser multi-object ifs l'adaptation en temps réel, en tenant compte des états cognitif s via signaux physiologiques ? Défis principaux : Mesure non intrusi ve de signaux physiologiques en milieu naturel (Darvishi et al., 2022 ). Identification d'états cognitifs à partir de données multimodal es (Azevedo et al., 2022). Adaptation dynamique de contenus/feedback (Mandouit & Hattie, 2023). Évaluation d'impacts sur apprentissage et bien-être (Ahmad et al., 2024 ; Upsher et al., 2022). Approche prop osée : Protocole expérimental pour collecte longitudinale non intru sive ; analyse avancée (psychologie + ML) pour patterns cognitifs ; modèles multi-objectifs (algorithmes génétiques, apprentissage par renforcement) ; validation via essais contrôle/test." - 15h-1 5h30 : Nour Bouchouchi (LFI) Title: Encoded and Expressed Gender Bias in LLMs: A Joint Study Abstract: During training, large language mo dels (LLMs) learn not only factual knowledge but also social regulari ties that can lead to gender bias in real-world applications. Most mi tigation efforts focus on reducing bias in generated outputs, typical ly evaluated using structured benchmarks. This raises two issues: out put-level evaluation does not reveal whether alignment alters the mod el’s underlying representations, and structured benchmarks may not reflect realistic usage scenarios. A unified framework is proposed to jointly analyze internal and expressed biases in LLMs. The results s how that, although alignment reduces visible bias, stereotypical asso ciations persist in internal representations and can be reactivated b y simple adversarial prompts. Furthermore, the effects of alignment d o not always generalize to more realistic settings, such as story gen eration. - 15h30-16h : Tristan Bersoux (SMA) Title: A Systemic Multi-Agent Approach for Analyzing Energy Policies Abstract: I will present my current work on an extension of TerraSim, an agent-based m odel that simulates the interactions between human activities (both i ndividual and collective), economic dynamics at the level of househol ds and firms, and their environmental consequences, including but not limited to land use change and greenhouse gas emissions. My contrib ution extends TerraSim with a detailed sub-model of the energy sector , enabling the evaluation of energy policies at both macro and micro levels. After a quick introduction to agent-based simulation, I will outline the current challenges of energy policy design and show how o ur model helps address them. I will then focus on our implementation method and present some early results, with a particular emphasis on the French electricity mix. - 16h10-16h40 : Louis Milhaud (CN) Title: Cut-based attacks on street networks Abstract: Assessing the robustness of a network generally involves observing the number of no des that remain interconnected when nodes or links are removed iterat ively. State of the art attacks, for example, target links present in many shortest paths or randomly remove elements, simulating failures . We focused on the case of urban networks where links represent stre ets and nodes represent intersections. Thus, a blockade (in the conte xt of a social movement) is modelled by the removal of links. On thes e networks, we applied efficient graph-partitioning heuristics to cre ate cutting-based attacks, and to empirically study the robustness of urban networks against such attacks - 16h40-17h10 : Nour BenAli (BD) Title: Effective Generation of Synthetic JSON Data Using Large L anguage Models Abstract: Automatic generation of synthetic data is essential for applications such as testing data pipelines, privacy-pr eserving data sharing, machine learning training, and benchmarking No SQL systems. In these contexts, data is commonly represented in JSON and constrained by JSON Schema. This work focuses on generating synth etic JSON data that respects schema constraints while remaining reali stic. Existing approaches often struggle to balance structural validi ty, semantic quality, and complex constraints. The research explores the use of Large Language Models combined with constraint-aware techn iques to generate JSON data that is both schema-compliant and semanti cally meaningful. - 17h10-17h40: Christos Malogiannis (CIAN) Tit le: Training event-driven neuromorphic systems Abstract: Deploying s piking neural networks on FPGA-based accelerators requires training s oftware that respects the operational constraints of the target hardw are. In this talk, I present a software framework for training event- driven spiking neural networks for an SNN accelerator developed in ou r lab. The framework preserves the hardware-oriented forward model du ring training, while using surrogate gradients for optimization and s upporting both post-training quantization and quantization-aware trai ning. I evaluate it on neuromorphic benchmarks including Card Symbols and N-MNIST, showing that models trained in software can be exported to FPGA-compatible weight formats and transferred to a hardware-orie nted inference flow with strong classification performance. - 17h4 0 : goûter DTSTAMP:20260603T061238Z DTSTART;TZID=Europe/Paris:20260605T133000 DURATION:PT2H URL;VALUE=URI:https://www.lip6.fr/liens/organise-fiche.php?ident=O1245 UID:LIP6/SEM/O1245 LOCATION:Campus Pierre et Marie Curie, LIP6, salle 105 25-26 GEO:48.847047;2.354619 END:VEVENT END:VCALENDAR