The Research On Algorithm Optimization Of Convolutional Neural Network Model Compression

Due to its complex network structure and a large number of network parameters,Convolutional Neural Networks require a large amount of computation and storage space,making it difficult for existing Convolutional Neural Networks to deploy applications on hardware devices with limited resources.Therefore,how to ensure that the accuracy of the Convolutional Neural Network falls within the allowable range,and to reduce the requirements for the computing capacity of hardware devices and the overhead of storage resources by streamlining the Nonvolutional Neural Network model,is an important research question.This paper mainly studies the compression of Convolutional Neural Network models.The specific work is as follows:1.Propose a double pruning algorithm based on the optimal ThresholdNetwork pruning is a technique for removing redundant parameters from the network to compress the Convolutional Neural Network model.The existing network pruning methods are insufficient in determining the network pruning Threshold and judging the importance of nodes.In order to solve these problems,this paper proposes a double pruning algorithm based on the optimal Threshold,which uses the sensitivity and correlation of the nodes as the basis for judging the importance.This algorithm can automatically select the optimal pruning Threshold that can balance the maximum sparse rate and the minimum error to prun the network,so as to achieve the purpose of model compression.Experiments show that the algorithm can effectively reduce network parameters and reduce running time.2.A R-Dropout regularization algorithm based on correlation dropDropout is a regularization method that prevents the model from over fitting by randomly deleting some nodes in the original network structure.As the size and complexity of Convolutional Neural Networks continue to increase,the number of linear correlation nodes in the network will increase.Related nodes will not only form data redundancy,cause waste of resources,and affect the efficiency of algorithm execution.Based on this,this article incorporates the idea of correlation into the Dropout regularization method and proposes an R-Dropout regularization method.This method effectively limits the number of parameters by deleting some highly correlated nodes with a certain probability.Experiments show that thealgorithm effectively improves the training convergence speed.Compared with the original Dropout algorithm,the algorithm performs more efficiently under the premise of the same number of iterations.3.A parameter quantization method based on binary K-means clusteringThe binary K-means algorithm makes up for the shortcomings of the initial clustering center randomly selected by the traditional K-means algorithm.This algorithm is strict in the selection of the initial clustering center,which makes the distance of each clustering center point far.This algorithm prevents the initial clustering center from being assigned to a cluster,reducing the possibility of the algorithm falling into a local optimum.First,the binary weighted K-means clustering algorithm is used to cluster the weights of each layer of the neural network,and then the original weights are replaced with the obtained cluster centers.In this way,weight sharing is achieved,and multiple connection weights in this layer share the same weight.In this way,the value of K is much smaller than the number of weights,which achieves the purpose of compression.