< Back to previous page

Project

Intermittent computing for the extreme edge in beyond 5G networks

Extreme edge computing assumes that significant processing can run on the sensors. Ideally, extreme edge processing can exploit the energy harvested from the environment, which means that energy supply is irregular. Intermittent computing divides complex tasks in small steps that can be executed when there is just enough energy available. Intermediate results are then saved efficiently so that the sensors can go into a deep sleep state. When new energy is available, the sensor can wake-up and resume the processing. Interesting platforms are being proposed in the state of the art for adding deep learning acceleration on the sensors, such as Google’s coral platform. Objectives are: (1) To study how much of the anomaly detection pipeline can run locally, and how many/much features/data should be communicated to the cloud for further processing there. (2) To co-design the local algorithms for various deep learning architectures, to make sure the algorithms match the chosen architectures optimally. A logical step is here for instance the use of quantized models. Beyond that, there also exist interesting approaches to achieve structured sparsity, and hence simpler models. (3) To divide the models into small parts that can be executed intermittently, while trading off memory access cost likelihood of losing intermediate results.

Date:21 Dec 2021 →  Today
Keywords:beyond 5G, Intermittent Computing, 5G networks, Edge Computing, Deep Learning, Machine Learning
Disciplines:Machine learning and decision making, Networking and communications, Wireless communications, Telecommunication and remote sensing, Computer communication networks
Project type:PhD project