Inference Acceleration Model of Branched Neural Network Based on Distributed Deployment in Fog Computing

摘要

Research based on deep neural networks (DNN) is becoming more common. In order to solve the problem that DNN needs to consume a lot of performance during the use prediction process and generate unacceptable delays for users, a distributed neural network deployment model based on fog computing is proposed. The distributed deployment of deep neural networks in fog computing scenarios is analyzed. A deployment algorithm based on Solution Space Tree Pruning (SSTP) is designed, and a suitable fog computing node deployment model is selected to reduce the delay of prediction tasks. An algorithm for Maximizing Accuracy based on Guaranteed Latency (MAL) is designed and implemented, and suitable fog computing nodes are selected for different tasks to exit the prediction task. Simulation experiment results show that compared with the method of deploying neural network models in the cloud, the model prediction delay of the distributed neural network model based on fog computing is reduced by an average of 44.79%. Reduced the average computing acceleration framework of similar algorithms by 28.75%.