Cloud-edge-end integrated Artificial intelligence based on ensemble learning

Abstract

Deep neural networks (DNNs) have been extensively used in the domains of artificial intelligence (AI) applications. Their inherent complexity primarily drives the deployment of DNN models in cloud environments. However, the geographical distance between the cloud and the end-users fails to meet the low-latency requirements of time-sensitive applications. Edge computing has emerged as a viable way to address this issue, nevertheless, the inherent constraints of limited resources on edge servers pose challenges in supporting intricate models. Solutions relying on network compression or model segmentation often fall short in meeting both performance and reliability needs. For the few ensemble-based solutions, the diversity between base models is not fully explored, and the low-latency advantage of edge computing is not fully utilized. In this paper, we propose a cloud–edge-end integrated approach for building an efficient and reliable DNN inference platform based on ensemble learning. In this design, heterogeneous models are trained on the cloud according to the resource constraints of edge servers, and the inference process is performed independently on each edge server, whose outputs are combined at the end-user side to get the final result. Furthermore, a diversity-based deployment scheme is proposed to build a user-centric network for edge AI. The generation of base models is explored, and the effectiveness of the proposed approach is demonstrated through two case studies.