Neural Network Realization of Support Vector Methods for Pattern Classification

I t is well known that the real-time processing ability of neural networks is one of their most important advantages. It allows neural networks to find numerous applications in many fields. However, the quadratic programming problem for training support vector machines (SVMs) is a computational burden [2,4] even though the training problem is formulated as a convex optimization problem. In particular, for the case where the matrix included in the quadratic program is semidefinite and training dataset is very large, the training of SVM is very time-consuming [5,6]. In this frequently happened case, more efficient algorithm has to be developed to speed up the training and evaluation of SVMs. In this respect, neural networks offer a very promising tool. If we combine the merits of the SVMs and neural networks, we can obtain a new neural network SVM with a better performance. Motivated by this idea, we apply a recurrent neural network to SVM training for pattern recognition. Specifically, a primal-dual neural network is exploited to solve the quadratic programming problem encountered in training SVMs. The properties of the network allow us to design SVMs without adjustable network parameters and give a better solution for ill-posed problems, e.g., semi-definite quadratic programming problem.