Potential methods in pattern recognition

Near-infrared libraries in the pharmaceutical industry: a solution for identity confirmation

“…The most commonly used among them are discriminant analysis (DA), 20 the k-nearest neighbour (KNN) function 21,22 and potential function methods (PFMs). 23,24 Modelling methods create volumes in the pattern space that possess different bounds for each class. Such bounds can be established in the form of correlation coefficients, distances (whether Euclidean, as in the PRIMA method, 25 or of the Mahalanobis type, as in the UNEQ method 26 ), the residual variance 27,28 or supervised artificial neural networks such as the multi-layer perceptron (MLP).…”

Neural Networks and Related Methods for Classification

“…However, we could use the analogous density estimation techniques to estimate p(x Iy) and to compute the discriminant from 'Trlp(xI1)/'Trop(xI0). This approach has been considered in monographs by Hand (1982) and Coomans and Broeckaert (1986). Estimating densities in high dimensions is notoriously difficult, and to use them for discrimination we are interested in regions where the densities of the two classes are comparable, usually in the tails of each.…”

Chemometrics

“…However, the authors caution that the probabilistic KNN procedure is very approximate in its present form (PR8). Reduction of the training set to generate a condensed nearest neighbor version of the probabilistic technique was also explored (PR10). Willet (PR41) has reported several algorithms for efficiently accomplishing condensed nearest neighbors searching based on the conventional (non-probabilistic) voting rules.…”