Design of an elliptical neural network with application to degraded character classification
Michael C. Moed, C.-P. Lee · 2002
A neural network architecture that is used to classify a set of patterns into one of a set of known classes is described. The network is comprised of a set of trainable neural processing units (neurons) that have an elliptical activation function and a set of adaptable connections. A fast training algorithm is provided for the network, which guarantees that all elements of an arbitrary training set can be correctly learned by the network in finite time. To demonstrate the network's ability to train, and its ability to quickly generalize and classify noisy test data, a network is developed to classify degraded omnifont alphanumeric machine printed characters. Using a training set of over 69000 characters and a separate test set of over 36000 characters, a classification accuracy of 97.5% with an average network throughput of 211 characters per second is achieved.>