A k-Winner-Takes-All Classifier for Structured Data

We propose a k-winner-takes-all (KWTA) classifier for structures represented by graphs. The KWTA classifier is a neural network implementation of the k-nearest neighbor (KNN) rule. The commonly used comparator for identifying the k nearest neighbors of a given input structure is replaced by an inhibitory winner-takes-all network for k-maximum selection. Due to the principle elimination of competition the KWTA classifier circumvents the problem of determining computational intensive structural similarities between a given input structure and several model structures. In experiments on handwritten digits we compare the performance of the self-organizing KWTA classifier with the canonical KNN classifier, which uses a supervising comparator.