Implementing the k-nearest neighbour rule via a neural network
Yan Qiu Chen, Mark S. Nixon, R.I. Damper · 2002
Presents a novel neural-network architecture which implements the k-nearest neighbour rule of pattern recognition. The architecture is synchronous (i.e. clocked) and has an essentially feedforward structure, but also incorporates feedback to control sequential selection of the k neighbours. Network training uses non-iterative weight calculations rather than iterative backpropagation. Analysis of the network shows that it will converge to the desired solution (classifying the input pattern according to the k-nearest neighbour rule) within 2 k clock cycles. The space complexity of the network is O(N/sub T//sup 2/), where N/sub T/ is the number of training patterns. This work offers prospects for an ultra-fast, parallel implementation of a proven pattern classifier.