Voici les orateurs prévus et le programme pour la journée du 5 juin (15-20mn de présentation + 10mn de questions) :

- 13h30-14h : Ioannis Kaskampas (CIAN)

Title: Federated Learning on Neuromorphic Nodes

Abstract: Federated learning (FL) allows multiple clients 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 devices because they communicate via discrete spikes rather than dense activations, and dedicated neuromorphic chips like Intel Loihi can execute them at a fraction of the power of conventional GPUs. Most existing studies, focus on a single FL paradigm and a single SNN implementation. This work studies federated learning on neuromorphic nodes across several FL paradigms — synchronous, hierarchical, asynchronous, and vertical — and across heterogeneous combinations of SNN backends and architectures. We evaluate the resulting trade-offs in accuracy, 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'apprentissage profond appliqué aux séries temporelles continue d'évoluer, et un sujet de recherche particulièrement actif actuellement est la conception, entraînement et utilisation de modèles profonds et préentraînés sur des jeux de données de grande taille pour résoudre des tâches sur d'autres jeux de données sans entraînement ou adaptation préalables. Pourtant, il y a toujours beaucoup de questions ouvertes, et certains modèles ne génèrent que des prédictions ponctuelles. Dans cette présentation, nous définissons un cadre théorique pour la prédiction des séries temporelles, en donnant des propriétés sur les fonctions utilisées pour évaluer la quantification de l'incertitude, et nous étudions deux modèles de séries temporelles récents, en comparant les intervalles qu'ils génèrent avec ceux générés par des techniques de prédiction conforme. - 14h30-15h : Aya Tounsi (MOCAH) Titre : Optimisation multi-objectifs visant à concilier l'état physiologique et les résultats d'apprentissage dans l'enseignement supérieur de longue durée

Résumé : "L'apprentissage individuel reste un défi malgré l'abondance de contenus pédagogiques : comment adapter ressources et feedback pour maximiser l'engagement, minimiser l'abandon et éviter surcharge cognitive ou ennui (état de flux, Csikszentmihalyi, 2014) ? Cette thèse intègre exploration de données éducationnelles, psychologie et physiologie pour développer des algorithmes IA dans un système de gestion de l'apprentissage. Question centrale : comment optimiser multi-objectifs l'adaptation en temps réel, en tenant compte des états cognitifs via signaux physiologiques ? Défis principaux : Mesure non intrusive de signaux physiologiques en milieu naturel (Darvishi et al., 2022). Identification d'états cognitifs à partir de données multimodales (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 proposée : Protocole expérimental pour collecte longitudinale non intrusive ; analyse avancée (psychologie + ML) pour patterns cognitifs ; modèles multi-objectifs (algorithmes génétiques, apprentissage par renforcement) ; validation via essais contrôle/test." - 15h-15h30 : Nour Bouchouchi (LFI) Title: Encoded and Expressed Gender Bias in LLMs: A Joint Study

Abstract: During training, large language models (LLMs) learn not only factual knowledge but also social regularities that can lead to gender bias in real-world applications. Most mitigation efforts focus on reducing bias in generated outputs, typically evaluated using structured benchmarks. This raises two issues: output-level evaluation does not reveal whether alignment alters the model’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 show that, although alignment reduces visible bias, stereotypical associations persist in internal representations and can be reactivated by simple adversarial prompts. Furthermore, the effects of alignment do not always generalize to more realistic settings, such as story generation. - 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 model that simulates the interactions between human activities (both individual and collective), economic dynamics at the level of households and firms, and their environmental consequences, including but not limited to land use change and greenhouse gas emissions. My contribution 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 our 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 nodes that remain interconnected when nodes or links are removed iteratively. 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 streets and nodes represent intersections. Thus, a blockade (in the context of a social movement) is modelled by the removal of links. On these networks, we applied efficient graph-partitioning heuristics to create 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 Language Models Abstract: Automatic generation of synthetic data is essential for applications such as testing data pipelines, privacy-preserving data sharing, machine learning training, and benchmarking NoSQL systems. In these contexts, data is commonly represented in JSON and constrained by JSON Schema. This work focuses on generating synthetic JSON data that respects schema constraints while remaining realistic. Existing approaches often struggle to balance structural validity, semantic quality, and complex constraints. The research explores the use of Large Language Models combined with constraint-aware techniques to generate JSON data that is both schema-compliant and semantically meaningful. - 17h10-17h40: Christos Malogiannis (CIAN) Title: Training event-driven neuromorphic systems

Abstract: Deploying spiking neural networks on FPGA-based accelerators requires training software that respects the operational constraints of the target hardware. In this talk, I present a software framework for training event-driven spiking neural networks for an SNN accelerator developed in our lab. The framework preserves the hardware-oriented forward model during training, while using surrogate gradients for optimization and supporting both post-training quantization and quantization-aware training. 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-oriented inference flow with strong classification performance.

- 17h40 : goûter