An analysis of discriminatory mechanisms in frequency-weighted memory array pattern classifiers

摘要

Cellular logic arrays for pattern classification may be configured in a variety of ways. Although a common configuration is to utilise simply binary weighting of feature samples occurring in a data training set, performance can generally be improved if a more sensitive frequency-of-feature-occurrence weighting scheme is adopted. In this form, a practical problem which must be addressed by the system designer is that of determining an appropriate weight value for features occurring with zero frequency in the training set.This paper discusses the difficulties encountered in attempting to make a global selection of the zero frequency component in discriminant function computation and, by examining the processing mechanisms of the classifier, points to the characteristics of data set composition which might allow an effective choice of this parameter on a more local basis. This is found to offer practical advantages in structuring a classifier, particularly since the optimisation procedures adopted appear to be directly compatible with an established multi-level classifier architecture.