Research And Application Of Deep Learning's Parallel Acceleration

With the continuous development of society and the advancement of science and technology,the core hardware equipment of computers has been gradually upgraded.That has greatly improved the performance of computers.Then it provides a better operating environment for deep learning to run,so that deep learning get a leap forward.Deep learning has made breakthroughs in the fields of computer vision,machine translation,etc.However,in scientific research and practical applications,appropriate hyperparameter is required to improve the model's performance.And faster training and inference speed is needed to reduce the waiting time.Therefore,this paper proposes the automatic configuration method of the model's hyperparameters.And we increase the speed of the data's regression prediction,which can improve the training and inference speed of the model to reduce time consumption.This paper first realizes the automatic configuration of the network model's hyperparameters based on parallel genetic algorithm.The method divides the individuals in the population into multiple subgroups according to the number of processes.Then each subgroup runs the genetic algorithm independently and exchanges some individuals to get the best individual and hyperparameters of the model.The method can enrich the diversity of the population,solve the problem of premature convergence and slow speed in genetic algorithm.So that it can improve the overall performance of the model and increas e the speed of hyperparameters' configuration.Then,we analyze the time series data and the spatial-temporal data of Internet of Things.And we also increase the speed of training and inference for them.In order to get the faster training and inference speed,this paper realizes the regression prediction of the data based on Simple Recurrent Unit.What's more,based on Simple Recurrent Unit,the paper considers the impact of the external factors on the regression prediction of time series.The external features of data are extracted by Convolutional Neural Network to remove the redundant features.During the prediction of spatial-temporal data,the paper considers data's periodicity and the shifting of periodicity.It solves the problem of periodic shifting,combines data's periodicity and attention mechanism to predict the data.At last,regression prediction network structures for time series data and spatial-temporal data are built respectively,which make the prediction performance of the network architecture better,and also increase the training and inference speed of the model.Finally,based on the above-mentioned analysis of time series data and spatialtemporal data,the article proposes a new application scenario based on the wireless sensor network.This method stores some information of packets in bloom filter,during the packets are transmitted.Then it detects the data received on edge server based on naive Gauss Bayes.If the data is abnormal,it will trace and change the path of the abnormal data packets.So that it can solve the data security problem of wireless sensor network during data transmission.