Un sistema integrato che raccoglie le informazioni da tutto l'Ateneo e permette di esplorare con rapidità, efficacia e trasparenza le competenze, mettendo a disposizione i dati relativi alla produzione scientifica, ai progetti di ricerca, ai brevetti e a tutte le informazioni di rilievo sull’attività di ricerca. Tramite la funzione di ricerca si possono scoprire gli ambiti di ricerca, le persone, le strutture e i risultati dalla ricerca condotta al Politecnico di Torino. È possibile cercare per parola chiave, nome persona, titolo progetto, in lingua italiana o inglese. Il sistema restituirà gli elementi presenti nei database del Politecnico di Torino, compresi i profili completi del personale.
EDGEML - EDGELM: Efficient and Privacy-Preserving Edge Deployment of Multimodal Large Models
Durata:
Responsabile scientifico:
Tipo di progetto:
Ente finanziatore:
Codice identificativo progetto:
Ruolo PoliTo:
Abstract
Large models (LMs) powering generative Artificial Intelligence (AI) services are typically hosted in centralized cloud infrastructures. This paradigm introduces significant challenges related to data sovereignty, privacy protection, energy consumption, and transmission latency, severely restricting their suitability for real-time, privacy-sensitive applications. Despite recent advances, current solutions for LM compression and edge deployment remain insufficiently adaptive to heterogeneous hardware distributions, diverse multimodal model requirements, and practical deployment constraints. The EDGELM project directly addresses these gaps by developing a context-aware, AI-driven framework that tailors LM compression strategies to heterogeneous edge-resource distributions, modality-specific compression characteristics, and application-specific accuracy constraints. Complementing this framework, an adaptive reinforcement-learning-assisted orchestration solution will be designed to optimize model cooperation and resource allocation, thereby improving inference latency and energy efficiency.
Furthermore, EDGELM will integrate hybrid quantum-classical fine-tuning methods and federated learning techniques to augment model expressiveness and training efficiency within limited resource budgets. EDGELM’s practical effectiveness will be validated through real-world deployment in a privacy-sensitive healthcare scenario, ensuring compliance with data protection and ethical standards, while evaluating user acceptance and trust with expert psychological insights. Supported by advanced edge-computing and quantum computing facilities at Politecnico di Torino, industrial expertise from Nextworks, and clinical input from healthcare partners, EDGELM will establish a robust methodology and practical roadmap for broader deployment of multimodal LMs at the network edge. This initiative supports Europe's transition toward privacy-compliant, sustainable, and efficient generative AI services.
Strutture coinvolte
Parole chiave
Settori ERC
Obiettivi di Sviluppo Sostenibile (Sustainable Development Goals)
Budget
| Costo totale progetto: | € 193.643,28 |
|---|---|
| Contributo totale progetto: | € 193.643,28 |
| Costo totale PoliTo: | € 193.643,28 |
| Contributo PoliTo: | € 193.643,28 |