Edge DEvice Training TEstbed
Large Language Models (LLMs) have demonstrated remarkable performance across various natural language processing tasks. However, deploying them on embedded platforms remains a significant challenge due to hardware constraints and limited energy availability. This project addresses the growing demand for secure, autonomous AI capabilities in edge computing environments—where reliance on cloud infrastructure is often impractical due to privacy concerns, latency requirements, or unreliable connectivity. Our objective is to evaluate, adapt, and implement a Small Language Model (SLM) optimized for resource-constrained embedded systems, with a strong emphasis on offline functionality, ultra-low power consumption, very low computational resources, and maintaining a responsive user experience. We specifically aim to implement this Small Language Model (SLM) within a smart ashtray application, designed to assist and support smokers in their smoking cessation journey by providing personalized advice and autonomous monitoring directly on the device.
Large Language Models (LLMs) have shown outstanding performance across natural language processing tasks. Yet, their deployment on embedded devices is hindered by limited memory, computing power, and energy efficiency. This project addresses these challenges by evaluating, adapting, and implementing a Small Language Model (SLM) designed for ultra-low power, resource-constrained systems. The proposed research will focus on: • Offline functionality, ensuring robustness even without network access. • Ultra-low energy consumption, suitable for battery-powered and energy-harvesting devices. • Reduced computational requirements, enabling deployment on platforms such as Raspberry Pi 5 or microcontroller-based systems. • Preserved user experience, maintaining responsiveness and relevance of interactions. The application case will be the Smokwit smart ashtray, an innovative device supporting smoking cessation in real-world environments. By embedding an autonomous conversational agent directly in the ashtray, smokers will receive personalized, privacy-preserving, and context-aware advice in line with evidence-based "5A" cessation protocols. Beyond this case, such optimization will pave the way for a wide range of applications where mobility, privacy, and autonomy are key. Expected outcomes include (i) a working prototype of a smart ashtray embedding an SLM, energy self-sufficient and delivering brief interventions offline, and (ii) a peer-reviewed scientific publication on design, implementation, and performance evaluation of such resource-efficient edge AI systems.
ARC4Health and Wellbeing