Training of a ML neural network for classification via recursive reduction of the class separation

A method for recursive training of a ML network for classification is proposed. It searches for the nonlinear discriminant functions corresponding to several small local minima of the objective function. The novelty of the proposed method lies in the transformation of the training data into new data with a deflated minimum of the objective function followed by iteration to obtain the next solution. It succeeded in finding solutions with lower misclassification errors than the solutions found after conventional training with random initialization of the weights for an OCR application