EC5: Edge–cloud collaborative computing framework with compressive communication

Abstract

With an increasing number of deep neural network (DNN)-based applications being deployed at the edges, edge–cloud collaborative computing has emerged as a promising solution to alleviate the burden on resource-constrained edges by collaborative inference. However, simply offloading part of DNN to the cloud introduces significant communication overhead during inference. In this paper, we propose EC5, an Edge–Cloud Collaborative Computing framework with Compressive Communication. The compression of the intermediate feature is formulated using information theory and jointly optimized with the DNN through end-to-end multi-task learning. By decomposing DNN parameters into a new space, EC5 enables efficient storage and update of models across various compression levels. An Adaptive Exit scheme is designed to retain high-confidence inputs on the edge for fast inference, reducing reliance on the cloud. Experimental comparisons with baseline methods prove that EC5 significantly conserves network bandwidth and reduces communication instances, with low latency and acceptable accuracy loss, showing flexibility across different communication scenarios.