Resource-Efficient DNN Training and Inference for Heterogeneous Edge Intelligence in 6G

Edge intelligence is expected to be a key enabler of the future sixth generation (6G) mobile network. However, the heterogeneous characteristics of edge intelligence, such as heterogeneous edge data, resources, and service requirements, pose challenges to deep neural network (DNN) training and inference at the edge. To fully tap the potential of DNN for 6G heterogeneous edge intelligence, a variety of solutions have been proposed to tackle the above challenges, such as federated learning and transfer learning, for DNN training; and DNN partitioning and early exiting, for DNN inference. In this paper, we provide a comprehensive survey about resource-efficient DNN training and inference in heterogeneous edge intelligence. We first discuss the challenges of DNN training and inference in heterogeneous edge intelligence. Then, we give a detailed review of the recent advances of DNN training and inference technologies in heterogeneous edge intelligence, including the principles and state-of-the-art solutions of these technologies. We further conduct a taxonomy and summary of the reviewed solutions to clarify their applicable scenarios and the limitations. Finally, we point out some potential future research opportunities on heterogeneous edge intelligence.