Classifying Glyphs: Combining Evolution and Learning

Abstract

This dissertation investigates the classification capabilities of artificial neural networks (ANNs). The goal is to generalize over the features of a writing system, and thus classify the writing system of a previously unseen glyph. The complexity of the problem necessitates a large network, which hampers the tuning of the weights. ANNs were created using three different hybrids of back-propagation (BP) learning and evolution, and a pure BP algorithm for comparison. The purpose was to find the method best suited for this kind of generalization and classification networks. The results suggest that ANNs are able to generalize enough to solve the classification task, but it is depending on the weight tuning algorithm. A pure BP algorithm is preferable to any of the hybrid algorithms, due to the size of the ANN. This algorithm had both the best classification results and the fastest runtime, in addition to the least complex implementation.