Antwerpen | More than two weeks ago
The swift growth of edge computing devices with low power consumption and advanced local data processing capabilities offers a promising avenue for implementing sophisticated machine learning models at the data origin. This study aims to harness the potential of deep learning, specifically focusing on Continual Learning for Large Language Models (LLMs), despite the limitations in computational power and storage. This approach allows edge devices to adapt to new data streams without losing previously learned knowledge, which is a vital feature of dynamic data settings.
The research will investigate techniques for the adaptation of large, cloud-based models to edge devices. This will be done both at the level of inference and fine-tuning. Edge-specific techniques that will be considered for investigation include model compression (quantization, pruning, knowledge distillation), fast decoding algorithms (early exiting, contextual sparsity), speculative decoding, attention optimizations (FlashAttention, sparse attention), system-level optimizations (operator fusion, memory management strategies, workload offloading), and hardware-level optimizations (specialized hardware accelerators like GPUs and TPUs). These optimizations are crucial for overcoming computational constraints while maintaining model performance.
The research will develop a deep learning framework tailored for ultra-low power edge devices, incorporating the Low Rank Adaptation (LoRA) algorithm and its friends. This enables adaptive learning with minimal alterations to model parameters, ensuring the framework remains compact and efficient. Additionally, the study will investigate data and model distillation techniques to address the computational constraints of edge devices while preserving crucial information.
Recognizing the sensitive nature of the data processed at the edge, this research emphasizes privacy protection. It will create mechanisms to secure user data during processing and when interacting with centralized systems, allowing edge devices to benefit from local data fine-tuning without compromising sensitive information confidentiality.
These privacy-preserving LLMs have the potential to transform various disciplines. For example, in healthcare, edge devices equipped with LLMs can analyze patient data in real time, providing immediate, personalized insights without external data transmission.
In summary, this research is anticipated to pave the way for more intelligent, autonomous, and secure applications of edge-based computing. This approach marks a significant step towards more responsive and smarter technological environments by enabling the continuous learning and adaptation of local data while maintaining privacy.
The research will address the following shortlist of research questions, which are to be further refined during the course of the PhD:
The integration of continual learning and edge-optimized large language models (LLMs) presents transformative opportunities across various industries. In semiconductor fabrication, edge devices with advanced AI capabilities can optimize production lines in real-time by continuously learning from sensor data, reducing defects, and improving yield without needing constant cloud-based intervention. In home automation, smart devices can offer more personalized and adaptive control, learning from user behavior and environmental changes to optimize energy consumption, security, and comfort, all while processing data locally to preserve privacy. In healthcare, wearable devices and diagnostic tools can provide personalized health monitoring and decision-making, analyzing patient data on the spot, thus enabling early detection and real-time recommendations without compromising sensitive information. Logistics can benefit from edge AI to optimize delivery routes, monitor inventory, and predict equipment maintenance needs, adapting to changing conditions dynamically and without reliance on constant cloud communication. These applications highlight the potential of edge-based LLMs to revolutionize industries by enabling smarter, faster, and more secure decision-making processes.
Required background: computer science, engineering
Type of work: 20% literature, 60% programming, 20% design space exploration
Supervisor: Steven Latré
Co-supervisor: Tanmoy Mukherjee
Daily advisor: Tanguy Coenen
The reference code for this position is 2025-098. Mention this reference code on your application form.