Energy Efficient Computation Offloading for DNN Tasks in Semantic-Assisted Mobile Edge Computing Systems

In this paper, we focus on semantic-aware computation offloading for executing the deep neural network (DNN) inference tasks in a mobile edge computing (MEC) system. To cope with the challenges of insufficient wireless resources during task offloading, we propose a compress-then-offload mode, in which each user first compresses its raw large-size task data into a small-size semantic feature and then uploads it to the edge server for remote execution. Specifically, we establish the relationship between the compression ratio and computation ratio for different DNN tasks. To achieve energy-efficient offloading, we formulate an optimization problem to minimize the energy consumption of all users by jointly optimizing the compression ratio, computation allocation, uploading time, and DNN layer selection. We propose a low-complexity iterative algorithm to solve the original non-convex problem. Simulation results demonstrate that our proposed semantic-aware computation offloading scheme can substantially reduce users’ energy consumption compared to the conventional offloading, especially with scarce wireless resources.