To emphasize class discrimination information, a feedforward network based on the feature distances between data is used. The motivation is that only information within a class boundary region (defined by the nearest neighbor distances) is involved in the training process. This is particularly useful in a highly non-linear or overlapping decision boundary case. In this paper we propose such a feedforward network, called a ck-nearest neighbor network, to analyze the k-nearest neighbor feature distances in relation to the discriminative distance of the classes. The k-nearest neighbor feature distances are weighted to give optimal class discrimination through training. This is analogous to learning the relationship between low-level feature information and the high-level class information. The method is demonstrated with the well known Fisher's iris data and the standard and noise induced interlocking spirals data.
class discrimination, feature distances, nearest neighbor, decision boundary, feedforward network