Research And Implementation Of SVM Based On Stochastic Computation

Support vector machine is one of the widely used classification algorithms.The learning and training processes have high complexity and long training time issues.Thus,the training task for a large amount of data cannot satisfy its application requirements very well.Stochastic computation,as a new number system,has the advantages of low complexity and strong fault-tolerant ability.We adopt the stochastic computation in the implementation of the support vector machines to significantly improve the issue for the application of support vector machines.Therefore,in this thesis,the stochastic computation is used in the support vector machines to reduce the computational complexity of support vector machines and speed up their training process.The main contribution of this thesis are as follows:Gaussian kernel function has a high proportion of computational complexity in the implementation of support vector machines.Stochastic computation is applied to the Gaussian kernel function calculation,in which,the Gaussian kernel function calculation structure(JEFA)based on stochastic computation is designed.This structure is realized by transforming the Gaussian kernel function,and deriving the forward transformation of the stochastic computation.According the testing results,we find that the mean square error of JEFA structure and traditional fixed-point method can be ignored when the length of the stochastic sequence is about 300;According the implementation results,we also find that the FF resource overhead of the JEFA structure is 1/5 of the traditional fixedpoint method,the LUT resource overhead is 1/20 of the fixed-point method,and the multipler resources are not required in JEFA.Therefore,the JEFA structure can significantly reduce the complexity without the computing performance losing compared with the traditional fixed-point scheme.Secondly,for the repeated calculation of the Gaussian kernel function set in the support vector machine,a kernel storage structure is proposed.Sequence minimum optimization algorithm,the specific implementation of the support vector machine,has repeated and cross-repetitive calculations in finding the working set and gradient.Therefore,a shared kernel function set is proposed to reduce the calculation of the kernel function set and the implementation complexity.Finally,the hardware implementation is proposed for the proposed support vector machine based on stochastic computation.It mainly includes data receiving and storage,main process control,kernel function set calculation based on stochastic computation,kernel storage and SVM training and learning modules.The main SVM training and learning modules include finding working set i,finding working set j,Lagrangian vector update,gradient vector update and other modules.The experimental data tests show that the support vector machine software simulation based on stochastic calculations(SVMSC)in terms of optimal parameters and classification accuracy is consistent with Libsvm,with an average speed of 1.5 times that of Libsvm.The hardware implementation of support vector machine with stochastic calculation(SVM-SC-H)can achieve the same classification performance as Libsvm whice the average speed is 40 times that of Libsvm.