Research And Implementation Of High Efficiency Support Vector Machine

SVM is one of the most important and classical statistical learning algorithms.It is mainly used in classification and regression fields.In the C-support vector classification application based on Gaussian kernel,the optimal training parameter Combination(C,?)plays a decisive role in the performance of the generated model.In practical applications,the SVM optimal training parameter Combination is usually searched by grid search method and cross-validation.However,the heavy computational load of this process takes too long a time with traditional search method,thus limiting the application of SVM in some occasions.This paper aims to improve the speed of SVM search for the optimal training parameter Combination and try to solve the time problem of LIBSVM,various methods are proposed in terms of algorithm and implementation.The main work of this paper is as follows:1.A shared dot product matrix(SDPM)algorithm is proposed.In order to solve the problem of dot product redundancy in the search of SVM optimal training parameter Combination by grid search method,the SDPM algorithm is proposed.This method firstly calculates the dot product of each sample point and the entire sample Combination and stores it as a dot product matrix.The method of uniformly reading from the dot product matrix in the subsequent search greatly reduces the process of searching for the optimal training parameter Combination.The calculation of the midpoint product.The theoretical derivation shows that under the specified search parameters,the amount of dot product to be calculated using the SDPM algorithm is only 1/60 of that of the traditional dot product calculation.2.The software implementation(SDPM-S)of searching the SVM optimal training parameter Combination based on the SDPM algorithm is completed.SDPM-S is applied to a specific Gaussian kernel-based C support vector classification problem and SDPM algorithm is introduced in its search process.Tests have shown that the search speed of SDPM-S is double the average LIBSVM search speed in the specified search mode.3.The software-hardware collaborative framework for searching SVM optimal training parameter Combination based on SDPM algorithm is proposed and implemented(SDPM-H&S).SDPM-H&S main structure is composed of upper computer and acceleration board.Among them,upper computer completes data classification and folding processing;accelerated board card field programmable gate array completes dot product matrix operation,kernel function operation and cross validation training.Accelerate the memory of the board to complete the dot product matrix storage,accelerate the PCI-e of the board to complete the data communication between the host computer and the acceleration board.Tests have shown that SDPM-H&S searches 30 times faster than LIBSVM in the specified search mode.