Partitioning Intelligent Applications Based Inception Networks in Mobile Edge Computing

Mobile terminals deploy deep neural networks (DNNs) in order to be able to perform tasks such as target detection and speech recognition. However, mobile terminals have limited computing power and storage space. It is difficult to run DNN-based applications directly on these devices. One solution is to offload some of partitioned tasks to the edge server for execution. Implementing this method is not easy due to two aspects: (1) Partitioning layer selection: a suitable partitioning layer needs to be selected in order to efficiently offload computational tasks to the edge servers. (2) Ineffectiveness of partitioning: the layer partitioning method does not significantly improve operational efficiency of DNNs in many cases, especially when dealing with complex structured DNNs such as Inception neural Network. This paper proposes a novel partitioning method, called INCEPTIONOFF. For a given DNN-based application, the method determines the partitioning layer, automatically determines the offloading scheme. We evaluated it on an image classification task using inception neural network and showed that this partitioning method achieves an average speedup of 205% and the fastest speedup of 263% compared to other methods.