Research On Structural Optimization Of Distributed Deep Neural Network Model

In recent years,deep learning has achieved great success in many applications.For example,face recognition technology based on deep learning has been widely used in our daily lives,greatly improving the safety and convenience of our lives.It can be predicted that with the acceleration of 5G commercialization,the number of Io T devices will increase dramatically,and the data generated by these devices will also explode.Distributed deep neural network can integrate distributed computing,deep neural networks,and big data well,so it has great potential.In order to improve the performance of the existing distributed deep neural network model,this thesis studies the structural optimization of the distributed deep neural network model.Since most Io T devices are embedded devices,their computing power,storage capacity,and electrical energy are generally limited.In order to make up for these limitations of Io T devices,this thesis improves the existing distributed deep neural network model and establishes a new distributed deep neural network model.Compared with the model before the improvement,the new model established reduces the size of the model distributed on the Io T device,which can reduce the computing power,storage space,and electrical energy required during the model inference process.Experimental results show that the improved model can achieve very high accuracy and speed up the inference of the model.Considering that each terminal device's perspective has different relative importance due to different positions,this thesis establishes a context-aware distributed deep neural network model based on this difference.Compared with the existing distributed deep neural network model,the biggest difference of the new model is the addition of a feature importance evaluation mechanism.When the importance of the feature is lower than the importance threshold,it will not be transmitted to the back-end for weighted fusion,which reduces the communication cost and energy consumption of the terminal device.The experimental results show that if the hyperparameters of the model can be adjusted well,the new model can reduce the communication cost of the terminal equipment under the condition that the model accuracy decreases little.