Edge Intelligence Based Collaborative Learning System for IoT Edge

Abstract

Edge Intelligence based collaborative learning systems have been developed to perform collaborative learning among multiple devices in a distributed environment. Majority of the collaborative learning systems have been designed using resources containing high computational power. It was identified that a system could be implemented to facilitate collaborative learning in resource constrained Internet of Things (IoT) devices. The existing collaborative learning systems were critically reviewed and analyzed to identify the ideal collaborative learning approach for resource constrained IoT edge. During the initial stages of the research, partitioned model training was identified as the most ideal approach. The research paved the way to design and implement two training architectures based on partitioned model training approach to facilitate environments with adequate and limited access to edge infrastructure. The proposed system utilized a hybrid deep learning model in partitioned model training approach for the first time. Furthermore, the research utilized a lightweight containerization mechanism to deploy the proposed collaborative learning system. The testing and evaluation phases of the research proved that the system was able to significantly reduce the resource consumption of the devices while achieving high model accuracy. The experimental setup reached up to 97% in model accuracy while consuming a significantly lower CPU consumption of 6.33%. The proposed system also proved to function efficiently by reducing energy consumption and reducing operational temperature by up to 4°C.