Evaluation of prototype learning algorithms for nearest-neighbor classifier in application to handwritten character recognition

The present work discusses what have been called 'imperfectly supervised situations': pattern recognition applications where the assumption of label correctness does not hold for all the elements of the training sample. A methodology for contending with these practical situations and to avoid their negative impact on the performance of supervised methods is presented. This methodology can be regarded as a cleaning process removing some suspicious instances of the training sample or correcting the class labels of some others while retaining them. It has been conceived for doing classification with the Nearest Neighbor rule, a supervised nonparametric classifier that combines conceptual simplicity and an asymptotic error rate bounded in terms of the optimal Bayes error. However, initial experiments concerning the learning phase of a Multilayer Perceptron (not reported in the present work) seem to indicate a broader applicability. Results with both simulated and real data sets are presented to support the methodology and to clarify the ideas behind it. Related works are briefly reviewed and some issues deserving further research are also exposed.

/pdf/decontamination-of-training-samples-for-supervised-pattern-1s1jgzlsm7.pdf

Decontamination of Training Samples for Supervised Pattern Recognition Methods

A class-dependent weighted dissimilarity measure for nearest neighbor classification problems

We have developed a reject option for VQ-based supervised Bayesian classification to improve classification accuracy by sieving out patterns that are classified with a low confidence value. A small codebook extracted from a learning vector quantizer (LVQ) is used to estimate the class-conditional densities of the feature vector. We adapt the two commonly used rejection criteria, outlier rejection and ambiguity rejection, for the VQ-based Bayesian classifiers. Using three high-level image classification problems, we demonstrate how local rejection criteria can improve the error vs. reject characteristics of our classifier over a global rejection method.

Yasuhiro Furukawa

Papers

Gradient descent learning of nearest neighbor classifiers with outlier rejection