Tag : Français

Analysis of hash tables with SIMD instructions

SIMD (Single Instruction, Multiple Data) instructions are a type of computer processor instruction that allows the same operation to be performed on multiple pieces of data at the same time. Typically, one can find all occurrences of a given byte value in a sequence of 16 bytes with a single instruction. When used effectively, this can greatly accelerate performance.

In this talk we explore some current SIMD versions of hash tables. For example, an idea is to build a table of size-16 buckets for the keys, each equipped with a 16-bytes word that contains a 1-byte fingerprint of the keys it contains: when looking for a key, one first checks with only one SIMD instruction if there are keys with the same fingerprint present in the bucket.

In a work that is currently in the writing phase, with Cyril Nicaud, we have successfully derived the mathematical equations that precisely describe the evolution of these tables as elements are inserted. Notably, we study the number of full-buckets (e.g., containing 16 keys). These solutions are not merely the functions resulting from the analysis of  standard hash tables multiplied by a constant factor. Instead, they are obtained through a probabilistic approach and are implicitly expressed as solutions of systems of differential equations. This relies on probabilistic techniques developed by Nicholas Wormald for studying dynamic random graphs.

Studies and modeling of information sharing and dissemination techniques in vehicular networks

This presentation explores vehicular ad hoc networks (VANETs) as a key component of intelligent transportation systems, enabling safer and more efficient road traffic through communication between vehicles and infrastructure. With the growing use of UAVs in urban environments, this work examines how vehicles and drones can cooperate to form a unified communication network. I introduce three routing approaches designed to improve inter‑vehicle communication in urban areas: a connectivity‑based routing method assisted by UAV supervision, a dual‑layer routing scheme separating ground and aerial communication, and a reactive routing strategy that predicts route expiration times. Simulation results highlight the strengths and limitations of these approaches and demonstrate their potential to enhance reliable end‑to‑end communication in VANETs.

Calcul efficace d’exécution d’automates

Dans cet exposé, je ferai une courte introduction de la vectorisation, une technique qui exploite des instructions matérielles spécifiques, les instructions SIMD, afin d’accélérer le traitement de donnée par le processeur. La vectorisation est le plus souvent réalisée automatiquement par les compilateurs, mais elle devient difficile lorsque le calcul est par nature séquentiel et que le contrôle de flot dépend fortement des données.

Un cas emblématique de programmes peu ou mal vectorisés automatiquement est celui de l’exécution d’automates. Dans cet exposé, je présenterai des méthodologies pour vectoriser les automates, qui s’appuient sur des résultats classique et récent de théorie algébrique des automates ainsi que quelques application concrètes de ces méthodologies.

Graphs of Shortest Paths

In this talk, we will explore graphs of shortest paths — directed acyclic graphs (DAGs) derived from shortest path traversals of a graph rooted at a fixed source. After establishing a precise definition of these structures, we will examine methods for uniformly sampling them and investigate their typical shape. Along the way, we will uncover interesting connections to bipartite graphs.

An Adaptable Digital Twin Architecture for Smart and Interoperable Electrical Roads

The Use of AI and Communication Networks for Early Alert and Prevention of Wildfires

This research develops an AI-based early wildfire detection system for edge device deployment in remote areas. By combining video sequence analysis with real-time environmental data (temperature, humidity, wind conditions), the approach targets early-stage smoke plume detection. The work involves implementing a multimodal deep learning architecture using a dataset of 50,000+ annotated images from 640 real wildfires across four countries. The work focuses on temporal pattern recognition to distinguish smoke from natural phenomena such as clouds and fog. The presentation will cover the research methodology, current implementation progress, and dataset characteristics. Future work will address model optimization for resource-constrained devices and communication network integration for real-time alert transmission.

Adaptive and Scalable Cluster Head Management for mMTC using Sampling-Based Agglomerative Clustering and Localized Reclustering

We present a lightweight and adaptive clustering framework for large-scale mMTC scenarios in 5G/6G networks. The approach combines low-complexity initial cluster formation, a drift-aware reclustering mechanism that reacts to spatial changes, and dynamic Cluster Head (CH) reelection based on residual energy, load, or time. A dedicated mobility handler ensures stable connectivity for moving nodes. Implemented in OMNeT++, the framework achieves significant improvements in execution time, network lifetime, delay, and energy efficiency compared to recent state-of-the-art protocols, making it highly suitable for scalable and resilient next-generation IoT deployments.

IoT, Wearables et IA Embarquée : Nouvelles Approches pour l’Interaction et la Mobilité Assistée

Cette présentation porte sur deux systèmes wearables innovants visant à améliorer l’interaction homme-machine et la mobilité assistée.
Le premier projet, Sense’Z, est un dispositif de reconnaissance de gestes de la main basé sur des capteurs inertiels et des modèles de machine learning embarqués sur microcontrôleur. Il permet une interaction fluide, intuitive et en temps réel dans un environnement IoT, offrant une nouvelle manière de contrôler des objets connectés sans contact. Les gestes reconnus peuvent être visualisés via une interface web/mobile.
Le second projet, EVAL Belt, est une ceinture intelligente destinée aux personnes malvoyantes. Équipée de capteurs de proximité et d’un système de retour vibratoire, elle détecte les obstacles, les passages étroits et intègre une fonction de détection de chute. Les événements (obstacles détectés, alertes de chute, trajets) sont enregistrés et peuvent être consultés en temps réel via la plateforme IoT (application/web) pour supervision et analyse.
Ces deux solutions illustrent le potentiel des technologies portables et connectées pour concevoir des interfaces accessibles, utiles au quotidien, et adaptées aux besoins réels des usagers.

Quantum computing and ZX-calculus

What if we could picture quantum computations not as daunting complex matrices, but as simple graphs of nodes and wires? The ZX-calculus offers exactly that: a rigorous graphical language where any quantum circuit can be represented as a network of connected nodes. In this interactive talk, I will introduce the basics of ZX-calculus in an accessible way. We’ll start with a quick refresher on quantum computing (just the basics, there is quantum background requirement for this talk), then see how these concepts translate into colorful ZX-diagrams. Using a handful of intuitive graph transformation rules, we will visually simplify and reason about quantum circuits — almost like solving a puzzle. For example, we’ll demonstrate how the famous quantum teleportation protocol, when depicted in ZX form, collapses into a trivially simple diagram, revealing its essence at a glance.

Along the way, a few fun quiz questions (via Kahoot!) will test and engage your intuition. Beyond the fun, we will discuss why this graphical approach is powerful for research. The ZX-calculus provides a sound framework for optimizing quantum programs and verifying algorithm correctness, and it even reveals surprising connections to combinatorics and algebra (its rewrite rules mirror an underlying Hopf algebra structure). I will also highlight current challenges and my ongoing work — in particular, extending these diagrammatic techniques to qudits (quantum systems with more than two levels), an open problem at the frontier of quantum computing theory. By the end of the talk, you’ll see quantum computing from a new angle and understand why one might say that « quantum computing is just graphs. »

View-based query processing

Are partial observations sufficient to retrieve information about the data that produced them? View-based query processing encompasses many reasoning problems that capture various aspects of this very natural question. This talk is an introduction to view-based query processing. We will define some of the most prominent tasks such as view consistency (how to check that the observation are coherent), certain answers (how to deduce reliable answers from partial data) and rewriting (how to rephrase desired information with respect to available data), and how they relate to each other. No prior knowledge required!