Multi-tier Edge Computing for Partitioned Neural Network

With the development of information technology, edge computing and artificial intelligence are gradually combined. At present, multi-tier edge computing network is often used in industry. The hierarchical structure of this edge network has something in common with that of neural network. Therefore, according to this characteristic, this paper proposes multi-tier edge computing for partitioned neural network. Specifically, in the work of this paper, based on the edge computing network, we first propose a kind of multi-tier collaborative computing of neural network. In this computing method, we can avoid large network transmission overhead and make up for cloud computing. Secondly, for our proposed computing method, we establish a time-delay optimization model, and design Ordered Elitist Genetic Algorithm (OESGA) based on Simple Genetic Algorithm (SGA) and Elitist Genetic Algorithm (EGA) to select the best partition solution for the neural network. Finally, we design experiments to demonstrate the advantages of our proposed computing method and algorithm. Firstly, we use LeNet5, AlexNet and VGG to test our computing method. By comparing the three neural networks with different computing method, we found that our proposed computing method can bring at least 50% delay improvement under the condition of limited bandwidth resources. Secondly, we build a 1000-1ayer neural network by randomly adding network layers. The experimental results show that with the increase of computing scale, our proposed computing method will gradually have more advantages. Finally, we compare the algorithm proposed in this paper with the traditional algorithm, and finally show that the algorithm proposed in this paper improves the convergence speed by about 4 times compared with the traditional algorithm.