Gaussian Synapse Networks for Handwritten Character Recognition

Abstract

In the context of improved higher order neural architectures for pattern recognition we have made use of a new type of higher order network containing gaussian synapses and developed the Gaussian Synapses Backpropagation Algorithm (GSBP) for the implementation of different types of pattern detectors. This paper concentrates on the presentation of the algorithm and its comparison to other structures in a typical benchmark problem i.e. the recognition of handwritten characters. The inclusion of gaussian functions in the synapses of the network allows the training algorithm to select the appropriate spatial information and filter out all that is irrelevant according to the training it has received. With this strategy it is possible to obtain very good recognition results with hardly any preprocessing and independently of backgrounds and slant using small networks that are quite easy to train.