A k-nearest neighbor artificial neural network classifier

The authors propose an artificial neural network architecture to implement the k-nearest neighbor (k-NN) classifier. This architecture employs a k-maximum network which has some advantages over the `winner-take-all´ type of networks and other techniques used to select the maximum input. This k-maximum network has fewer interconnections than other networks, and is able to select exactly k maximum inputs as long as its (k-1) ^th and k^th maximum inputs are distinct. The classification performance of the k-NN classifier is exactly the same as that of the traditional k-NN classifier. However, the parallelism of the network greatly reduces the computational requirement of the traditional k-NN classifier. Unlike the multilayer perceptrons which involve slowly converging back-propagation algorithms, the k-NN artificial neural network classifier does not need any training algorithm after the initial setting of the weights