A hybrid fast inference approach with distributed neural networks for edge computing enabled UAV swarm

Abstract

Nowadays, unmanned aerial vehicle (UAV) swarm supported by mobile edge computing is attracting more and more attention, such as smart agriculture, smart transportation, smart security surveillance, and smart environmental monitoring. Nevertheless, small UAV devices cannot deploy advanced best-performing DNN models with insufficient computing and storage resources. Aerial computing centers serving as edges can collect tasks from UAVs, reduce UAV load, and provide high-accuracy inference. Limited radio resources in offloading and load imbalance in computing centers could lead to significant latency in task inference. Considering image processing in UAV applications, we propose a hybrid inference framework combining early exit and task offloading based on distributed neural network. The bottleneck-designed DNN provides a smaller intermediate data size than the original input. By deploying shallow neural networks at the drone terminals and advanced DNNs at the aerial computing servers, latency-constrained tasks can be exited after local inference, offloaded to the edge after local computation, or directly offloaded to the edge. Compared to binary or partial offloading, the early-exit-based framework provides fast inference as the workflow stops when a confident result is obtained. Combining task offloading into the inference process offers a more flexible option to release the UAV computing load. The proposed approach efficiently utilizes wireless transmission and balances the computation load to maximize the system inference accuracy within the latency constraints. Practical task inference simulations demonstrate that our approach obtains higher system accuracy under different radio resources, task generation rates, and terminal computing capacities conditions.