Novel Neural Architectures for Recognition of Handwritten Characters

This paper presents an overview of novel networking strategies for neural networks which significantly improves the training and recognition performance of such networks whilst maintaining the generalisation capabilities achieved by existing architectures. A number of different architectures are introduced based on two major principles. The first of these employs RAM-based neurons arranged in multilayer clusters and the second involves modifying the existing weight structure of a back-propagation network to utilise weights taken from a given domain of Clifford algebra. The architectures are described in terms of the structure of the neurons they employ.