The DSFPN: a new neural network and circuit simulation for optical character recognition

A new type of neural network for recognition tasks is presented. The network, which is called the "dynamic supervised forward-propagation network" (DSFPN), is based on the forward only version of the counterpropagation network (CPN). The novel DSFPN is trained using a supervised algorithm and can grow dynamically during training, allowing allographs in the training data to be learned in an unsupervised manner. Training times are comparable with the CPN while giving better classification accuracies than the popular multilayer perceptron (MLP). Data preprocessed using Fourier descriptors show that, on average, the DSFPN is trained in 1353 times fewer presentations than the MLP networks and gives best recognition accuracy of 98.6%. Moreover, data preprocessed using wavelet multiresolution analysis gives a very high recognition accuracy; the best accuracy is 99.792%. Results show the effectiveness of the DSFPN and justify a hardware implementation to enable fast data classification. A circuit implementation for the DSFPN competitive middle layer is presented, and simulation results show that it can perform reliable pattern recognition at a rate of over 100 kHz.