The dynamic supervised forward-propagation neural network for handwritten character recognition using Fourier descriptors and incremental training

A new neural network, called the “dynamic supervised forward-propagation network”, is presented for recognition tasks. The network is based upon the counterpropagation network but trains using a supervised learning algorithm and incremental training. In addition it permits unsupervised dynamic growth of the middle layer allowing unknown subclasses to be learnt. The performance of the network, in classifying handwritten numerals presented as Fourier descriptors, is compared with the performance of other, neural networks: the back propagation network and the Counterpropagation Network. The new network shows an increase in classification accuracy over the back propagation network along with a 404 times decrease in the number of training pattern-presentations required. A considerably higher accuracy is obtained than that for the Counterpropagation network