An Attention-Enhanced Edge-Cloud Collaborative Framework for Multi-Task Application

Zhipeng Zhang, Wenting Ma, Feng Li, Renjie Tang, Jinlang Wang, Wai Chen

Abstract: We investigate in this paper the issue of Internet of Things (IoT) architecture to take advantage of edge-computing and intelligence. Edge devices are often shared by services packaged running different applications. This requires edge infrastructure to be able to switch between applications. Also, most IoT systems usually concurrently run multiple applications that target different tasks. To address these problems, we propose an attention-enhanced cloud-edge collaborative framework that is composed of one universal block (UB), a few initial layers of the deep model, located at the edge and multiple parallel task-specific attention blocks (TSABs) located in the cloud. For each task, a deep model is constructed by the UB and a TSAB where the UB is responsible for the general feature extraction and the TSABs are responsible for the subsequent task-specific processes. We empirically demonstrate the effectiveness of our approach with several state-of-the-art Convolutional Neural Network (CNN) models, including VGG and ResNet network architectures on various image classification datasets.