Semi-supervised Algorithms in Resource-constrained Edge Devices: An Overview and Experimental Comparison
Mahdi Barhoush, Ahmad Ayad, Anke Schmeink · 2022 IEEE International Conferences on Internet of Things (iThings) and IEEE Green Computing & Communications (GreenCom) and IEEE Cyber, Physical & Social Computing (CPSCom) and IEEE Smart Data (SmartData) and IEEE Congress on Cybermatics (Cybermatics) · 2022
Due to scalability and privacy concerns, machine learning computation is progressively moving from the cloud to edge devices where the data usually resides. However, the resources available to edge devices, such as memory, computing power, and latency, are limited. Many strategies in semi-supervised learning (SSL) help to train models with a small amount of labeled data and a greater number of unlabeled data that is easy to obtain and collect. However, these algorithms are typically compared in terms of their maximum accuracy performance, with little attention paid to how well they operate on devices with limited resources. This paper addresses this issue and gives an experimental analysis and comparison of modern semi-supervised learning algorithms when they are run on edge devices with constrained resources. The results suggest that simple SSL algorithms may become preferable to more complex ones for edge devices as they provide adequate performance with less computation power and memory usage.