An Adaptive Neural Architecture Search Design for Collaborative Edge-Cloud Computing

Abstract

With the development of deep neural networks, the smart edge-cloud scenario is expected to meet the diverse and personalized requirements of users. However, the exiting edge-cloud architecture requires the deployment of manually designed neural networks. Providing adaptive network architecture in this environment, especially for heterogeneous edge devices, becomes the major challenge. To this end, we design the customized network architecture with neural architecture search for adaptively neural network search. Specifically, our architecture can be identified in three parts: cloud layer, equipped with huge amounts of computing resources, conducts multi-objective architecture search on a proxy dataset; the edge layer aims to train the optimal architecture based on the target dataset from scratch, which makes full use of the computing resource of the edge server; and the user layer deploys the multi-objective model in diverse devices for inference in real time. Furthermore, we deploy an industrial sensor monitoring scenario as a case study to search for the temporal convolutional network, demonstrating the effectiveness of the proposed architecture. Experimental results show the feasibility of the proposed architecture.