Sigma-pi implementation of a nearest neighbor classifier
H.-C. Yau, MICHAEL T. MANRY · 1990
In practical pattern-recognition applications, the nearest-neighbor classifier (NNC) is often applied because of its near-optimal performance and because it does not require a priori knowledge of the joint probability density of the input feature vectors. However, the NNC has problems. For a small number of example vectors, it is difficult to optimize the NNC with respect to the training data. This problem is resolved by mapping the NNC to a sigma-pi neural network, to which it is partially isomorphic. A variation of backpropagation learning is then used to improve classifier performance. As an example, the approach is applied to the problem of hand-printed numeral recognition. Significant improvements in classification error percentage are observed for the training data and testing data