Research On Neural Networks For Support Vector Machines Learning And Hardware Implementation

Support vector machines (SVMs), based on structural risk minimization, have already been used widely as a kind of tool for classification and regression, because of its virtues as good generalization, low error, easy disposal on mathematics and simple structural explanation and so on. Up to now, the investigation on SVM mainly has concentrated on study of the theory and optimization of arithmetic. Comparing with it, the research on application and realization of arithmetic is less. By far, there have been only limited research reports about this field. And most of the arithmetic can only be realized by the software rather than by the analog hardware, which distinctly restricted the application of SVM.In many engineering applications, the demand on real-time data processing is often needed. Therefore, parallel and distributed approaches to training SVM are necessary. Considering the strong real-time processing ability of neural networks, we integrate neural network into SVM and use partially dual formulation to transform the quadratic programming problem into the dynamic equations of neural networks. The standard SVM learning neural network is constructed, whose topology adapt to analog circuits implementation easily. Such an idea not only can increase the speed of training SVM greatly, but also provides a new thought way for application of SVM. Based on it, this paper further proposes a learning neural network on least squares support vector machine. Compared with former, the proposed neural network avoids using projection variables and is trained using Lagrange multiplies directly, which eliminates the nonlinear parts of the neural network mentioned in SVM neural network. It results in a simpler and more efficient analog circuit realization in real-time application. Moreover, only the classifier learning network is discussed in former, but not refers to regression problem. The proposed network for LS-SVM learning can both solve classification and regression problems almost without changing its topology. And the stability of their corresponding neural networks also is proven in the paper. The simulation results based on Simulink and Pspice and the experiment of circuit implementation illustrate the effectiveness of proposed neural network.