CroApp: A CNN-Based Resource Optimization Approach in Edge Computing Environment

With the rapid growth of intelligent devices and the emergence of various DNN-based intelligent applications, cloud-only processing schemes can hardly guarantee the latency requirement and faces the risk of data leaking. Thus, recent work leverage on-device inference schemes to solve these problems through edge computing, while the most crucial issue in this scheme is how to deploy various computation-sensitive intelligent applications on the resource-constricted end device. In view of this challenge, we proposed a novel DNN-based resource optimization service (DROS) to adaptively optimize the memory and computation resources consumption of concurrent intelligent applications running on the end devices, and thereby enhance the scalability. In DROS, we adopt model pruning and computation sharing as model optimization approach to reduce the resource consumption as well as enhance the scalability. The experimental results show that DROS reduces the memory and computation resource reduction, scalability, and application performance.