Optimizing Task Completion Rate in Multi-user Edge Intelligence Networks Through Neural Network-Based Energy Management with Partitioning and Offloading

This paper addresses the challenge of running large-scale neural networks (NNs) directly on energy-constrained IoT devices, proposing the Energy-Management Neural Net Partitioning and Offloading (EMNPO) technique. Through NN partitioning, where specific layers are outsourced to edge servers, EMNPO utilizes dynamic programming and the theorem of minimum cut/maximum flow to efficiently solve the optimization problem posed by constrained server resources. Decomposing the problem into manageable sub-problems, EMNPO achieves solutions in polynomial time, significantly improving NN inference task completion rates compared to other methods. An investigation into the impact of energy constraints, NN types, and device counts demonstrates EMNPO’s superiority, highlighted by simulation results with real NN models. This affirms the practical effectiveness of EMNPO in enhancing the efficiency of edge intelligence systems in real-world scenarios, addressing the impracticality of running large-scale NNs directly on IoT devices with constrained energy.