Dynamic prototype addition in generalized learning vector quantization

Abstract

Learning Vector Quantization (LVQ) is a powerful supervised learning method for classification that uses a network of prototype vectors to form a deci- sion surface. Generalization theory shows there is a non-trivial number of prototype vectors that yield the best generalization. Although it is typical to assign the same number of prototype vectors for each class, other LVQ methods add prototypes dynamically (incrementally) during training. This work offers an extension to the existing dynamic LVQs that minimizes the cost function of Generalized LVQ by focusing on the set of misclassified samples. This cost minimization occurs between the largest cost-contributing class and its nearest “confuser class”. A comparison is made between other prototype insertion methods and compares their classification performance, the number of prototype resources required to obtain that accuracy, and the impact on the cost function.