A recurrent neural network-based method for training probabilistic Support Vector Machine

Abstract

In this paper, Support Vector Machine (SVM) is reformulated to a recurrent neural<br><br>network model which can be described by the nonlinear dynamic system. In the proposed<br><br>algorithm, an iterative training procedure is proposed independent of initial point. Also<br><br>probabilistic constraints are recommended for reducing effect of noisy samples in training<br><br>procedure and appearance of incorrect Support Vectors (SV). Probabilistic constraints admit<br><br>using knowledge about distribution function of samples. A set of differential equations is used<br><br>to modelling of the proposed probabilistic SVM. These equations are converged to optimal<br><br>solution for SVM. The Euler method is used to solve differential equation. The primal and<br><br>dual problem of SVM is solved by this model. Enough information is given for finding<br><br>optimal hyper plane. Capability of the proposed method is shown by experimental results in the<br><br>Optical Character Recognition (OCR) and synthetic data.