Description du projet
Une nouvelle piste pour l’IA sans fil
Les systèmes de communication sans fil sont confrontés au défi de devoir traiter toujours plus de données dans des environnements dynamiques. Les méthodes traditionnelles de conception des récepteurs peinent à suivre l’évolution rapide des canaux sans fil. Qui plus est, la puissance et les ressources informatiques des dispositifs sont limitées, ce qui complexifie le traitement de grandes quantités de données. Les solutions d’IA actuelles, qui s’appuient sur d’énormes réseaux pré-entraînés, sont mal adaptés à ces conditions. Le projet FLAIR, financé par le CER, entend résoudre ces problèmes en créant une nouvelle forme d’IA flexible, conçue spécifiquement pour les communications sans fil. Il concentre ses efforts sur la conception de récepteurs légers, l’apprentissage continu et l’utilisation efficace des données, proposant une approche plus adaptable et plus respectueuse des ressources.
Objectif
Artificial intelligence (AI) is envisioned to play a key role in future wireless technologies, with deep neural networks (DNNs) enabling digital receivers to learn to operate in challenging communication scenarios. However, wireless receiver design poses unique challenges that fundamentally differ from those encountered in traditional deep learning domains. The main challenges arise from the dynamic nature of wireless communications, which causes continual changes to the data distribution, combined with the limited power and computational resources of wireless devices. These challenges impair conventional AI based on offline trained massive DNNs. Our ambitious goal is to introduce a new form of flexible lightweight AI that is particularly tailored for wireless communications. Our approach is based on a holistically revisiting the three fundamental pillars of AI – the architecture, dictating the family of learned mappings; the training algorithm that tunes the architecture; and the data based on which learning is carried out. Accordingly, we focus on three objectives – 1) design trainable receiver architectures that are lightweight and support adaptation to rapid channel variations; 2) establish a new learning paradigm that deviates from conventional training, and is based on viewing continual learning as a dynamic system; and 3) propose techniques to accumulate online data sets that are sufficiently informative for learning purposes while being small enough not to induce notable complexity in training. This is a fundamental depart from conventional deep learning, based on highly-parameterized DNNs trained with massive data sets using lengthy learning procedures. Our preliminary data show that this paradigm shift achieves substantial performance, robustness, and complexity gains over conventional deep receivers. The project will transform how communications systems are studied, and profoundly impact a multitude of applications that rely on wireless communications.
Programme(s)
- HORIZON.1.1 - European Research Council (ERC) Main Programme
Thème(s)
Appel à propositions
(s’ouvre dans une nouvelle fenêtre) ERC-2024-STG
Voir d’autres projets de cet appelRégime de financement
HORIZON-ERC - HORIZON ERC GrantsInstitution d’accueil
84105 Beer Sheva
Israël