Improved generalization in ANNs via use of conceptual graphs: a character recognition task as an example case

The encoding schema used at the output of the ANN as a means for yielding improved generalization is modified. The encoding schema of assigning one element of the output layer to each category of the classification task is used as the reference for demonstrating improvement. The improved encoding schema is derived as follows. The human process of thinking about letters in terms of components (e.g., long lines, short lines, curves, etc.) in certain relationships (e.g., touch, abut, intersect, etc.) is modeled using the conceptual graph formalism, and then these are turned into codes usable by ANNs (C-R or concept-relation, vectors). There is significant improvement in the generalization performance of the ANN trained with this encoding schema vs. the base encoding schema