Active GP Classifiers for Handwritten Digit Recognition

Classification techniques strive to use features selectively to provide an optimal balance between classification accuracy and robustness. In traditional classification, all the classes in the n-class problem are equally important and the classification method tries to provide the best possible separation between all classes. Recent advances in classification termed active classification make selective decisions for specific subsets of classes within the multi-class problem. Typical examples of such scenarios are cases when there is a very high degree of confusion between two particular classes and a sharper solution is deemed useful. The design for active classifiers requires the underlying pattern recognition technique to blend feature discovery within the classifier training phase. When features are derived from handwritten character images, different levels of informative detail can be present in different regions of the image. Certain features are obviously more distinguishing than others when it comes to separating various subclasses in a multi-class problem at least for this domain. We conjecture that this dual task of feature selection and classifier training can be combined effectively using the Genetic Programming (GP) paradigm as described in this chapter. This chapter includes a comprehensive analysis of these and similar issues. Comparative results on a variety of datasets in handwritten digit recognition using a variety of parameters are also reported.