Neural Architecture Search for Computation Offloading of DNNs from Mobile Devices to the Edge Server

Abstract

With the rapid development of modern deep learn-ing technology, deep neural network (DNN)-based mobile appli-cations have also been considered for various areas. However,since mobile devices are not optimized to run the DNN appli-cations due to their limit of computational resources, severalcomputation offloading-based approaches have been introducedto overcome the issue; for DNN models, it was reported that,their elaborate partitioning, which allows that input samples arepartially executed on mobile devices and then the edge serverprocesses the rest of the execution, can be effective in improvingruntime performance. In addition, to improve communication-efficiency in the offloading scenario, there have been also studiesto reduce transmitted data from a mobile device and the edgeserver by leveraging model compression. However, the existingapproaches have the root limitation that the performance eventu-ally depend on that of the architecture of original DNN models.To overcome this, we propose a novel neural architecture search(NAS) method to consider the computation offloading cases. Onthe top of the existing NAS approaches, we additionally introduceresource and channel selection mask. The resource selection maskeffectively divides the operations in the target model into thosefor a mobile device and the edge server; the channel selectionmask allows to transmit only selected channels to the edge serverwithout the reduction of task performance (e.g., accuracy). Basedon the two additional masks, for the NAS procedure we introducea new loss function to take into account end-to-end inferencetime as well as the task performance which is the original goalof NAS. In the evaluation, the proposed method is compared toexisting approaches; we see from the experimental results thatour method outperforms both the previous NAS and pruning-based model partitioning approaches.