Research On Machine Learning Methods For Mobile Edge Computing

With the advent of the Internet of Everything era,the traditional cloud computing model has been unable to meet the rapid response to massive image recognition,resulting in a sharp decline in the quality of user experience.Therefore,based on edge computing technology,how to use image discriminative feature extraction methods to process massive image data at the edge of the network to provide users with low-latency and high-accuracy image recognition services has become a hotspot in academia and industry.This thesis mainly studies the image discriminative feature extraction method based on edge computing,considering how to extract the discriminative features of image data under the edge computing platform to provide users with low-latency and high-performance image recognition services.We have found that it is difficult to extract the discriminative features of data on terminal devices,it is difficult to extract the effective discriminative features of data on edge servers,and it is difficult to meet the various needs of users in different applications.In addition,this thesis also studies the problem of collaborative training of deep learning models in edge computing environments.The main innovations of this thesis are:1.Aiming at the problem that it is difficult to extract the discriminative features of the data on the mobile device based on the currently available resources,a feature extraction method for edge-end collaborate resource-aware is proposed.This method uses the data information on the edge server to assist in generating the discriminative feature extractor,which solves the problem that the existing method cannot extract the discriminative feature of the data on the mobile device.This method first proposes a discriminant feature extractor generation algorithm.Then,the method proposes a nested discriminant feature extraction algorithm to divide the extractor generated by the extractor generation algorithm into a multi-capacity discriminant feature extractor and deploy it on the terminal device.Finally,experiments verify that the method reduces network transmission traffic and extractor switching overhead while improving recognition accuracy.2.Aiming at the problem that it is difficult to extract the effective discriminative features of the data on the edge server,an effective discriminant feature extraction method for cloud-edge collaboration is proposed.This method generates a discriminative feature extractor through fine-grained analysis of the data structure information on the cloud server,which solves the problem that the existing method cannot extract the effective discriminative feature of the data on the edge server.This method first proposes a weight adaptive projection matrix learning algorithm.Then,the method sends the extractor generated by the projection matrix learning algorithm to the edge server to extract the discriminative features of the data on it.Finally,it is verified by experiments that the method reduces the network transmission traffic while improving the recognition accuracy.3.Aiming at the problem that it is difficult to meet the diverse needs of users for response time and accuracy,a hierarchical cloud-edge collaborative discriminant feature extraction method is proposed.By deploying different discriminative feature extractors on different levels of edge computing platforms,this method solves the problem that the existing methods cannot meet the diverse needs of users for response time and accuracy.This method first proposes a hierarchical optimization discriminant feature extractor generation algorithm.Then,the method proposes an extractor deployment algorithm to divide the extractor generated by the extractor generation algorithm into multiple sub-extractors and deploy them on the edge computing platform.Finally,experiments verify that the method can improve the recognition accuracy while meeting the diverse needs of users.4.Aiming at the problem that it is difficult to obtain high performance for simple deep learning models deployed on edge servers,a cloud-edge collaboration deep learning model training method is proposed.This method solves the problem that existing methods cannot obtain high-performance simple deep learning models by using complex deep learning models to assist in training simple deep learning models.This method first proposes a model initialization algorithm.Then,the method also proposes a model update method that uses unlabeled data and the collaboration of complex deep learning models and simple deep learning models to further improve the performance of the latter model.Finally,experiments have verified that this method can improve the performance of simple deep learning models.