| In recent years,with the development of deep learning,the models of deep neural networks have become increasingly complex and massive,and the required memory and input sample memory size have also increased.This not only reduces the training and inference speed of deep neural networks,but also limits their deployment on some Io T devices with small memory and poor computing power,It will also greatly increase the transmission volume and cost of samples collected by Io T devices and transmitted to central cloud nodes.The distributed deep neural network model is a deep neural network that uses edge computing technology.The distributed neural network deploys part of the deep neural network on terminal devices and edge devices to run using edge computing.The distributed neural network reduces the traffic between the cloud and the edge,terminal devices,and speeds up the reasoning speed.Moreover,the input samples processed by edge computing will not directly display the user’s information,which protects the user’s privacy.However,there are some issues with distributed deep neural network models.For example,terminal devices and edge devices have smaller memory,limited computing power,and higher transmission costs between terminals,edge devices,and cloud devices.In response to the above issues,the main work content of this thesis is as follows.1.This thesis improves the existing distributed deep neural network model and obtains an adaptive distributed deep neural network model.Compared with previous models,it simplifies the neural network module structure on terminal devices.After experimental verification,this adaptive distributed deep neural network model structure greatly improves the accuracy of terminal exports while simplifying the network structure of terminal devices.2.Due to the presence of multiple terminal devices in distributed deep neural networks,in image classification tasks,the input of each terminal device is a different part of an image or a different perspective of the same object.The influence of the features obtained by different terminal devices on the final classification results also varies,which means their importance is different.This thesis proposes a feature weighted fusion method based on likelihood estimation for distributed deep neural networks.This fusion method can calculate the importance of each feature through likelihood calculation and perform weighted fusion based on the importance of the features.Through experiments,it has been proven that using this feature fusion method can reduce communication traffic and costs while maintaining high accuracy by setting appropriate importance thresholds.3.Considering that the terminal part of distributed deep neural networks usually has a small memory capacity and is difficult to deploy large-scale neural networks,this thesis combines adaptive threshold binarization with binary neural networks to obtain a binary neural network based on adaptive threshold,and optimizes the model structure and training method of the binary neural network.Through experiments,it has been proven that the use of adaptive threshold binary neural networks in the terminal module of distributed deep neural networks greatly reduces the model memory and computational complexity while maintaining high accuracy. |
