Adaptive DNN Partitioning Strategy for Optimized User Fitness in Edge Computing Networks

Deep neural network (DNN) models are more frequently utilized in consumer electronics, posing challenges due to their intricate nature and substantial computational requirements for deployment on resource-limited devices. Direct deployment on devices necessitates significant computational resources, whereas cloud computing can mitigate this issue. However, extensive data transfer between the cloud and devices may result in elevated communication latency, hindering the fulfillment of computational needs in consumer electronics. To enhance DNN inference performance, this research investigates the optimization of DNN partitioning within a collaborative computing framework involving users, edge devices, and the cloud. A mathematical model is developed that considers latency, device energy consumption, and cloud rental cost preferences. The study proves that optimizing DNN inference based on user fitness is a challenging problem and introduces an adaptive partitioning strategy that adjusts to varying inference request loads. Additionally, an online optimization algorithm driven by a swarm is proposed to minimize system operating costs in dynamic settings. Experimental results demonstrate a 10.2% enhancement in user fitness compared to existing approaches, showcasing improved cost management without compromising model accuracy.