Feedforward neural network for handwritten character recognition

An analysis of feedforward neural networks for handwritten character recognition was performed to improve the learning capability and accuracy of classification, which are limiting factors of back-propagation. The authors describe two methods which attempt to tackle the shortcomings of back-propagation yet keep the feedforward organization of the neural network. These methods give results comparable to back-propagation, while requiring less training time and a simpler architecture. The first method rejects any pattern which differs from the training data more than a threshold, established during training. The second method involves clustering techniques selecting the most representative patterns as cluster centers. Both methods present the design of a neural network for handwritten digit recognition, and are based on the Parzen window estimates defining the vector space for different classes