Learning vector quantization with local subspace classifier

In this paper, generalized learning vector quantization (GLVQ) with local subspace classifier (LSC) is proposed for achieving high accuracy with a small memory requirement. In a training phase, the k-closest prototypes to an input training sample are moved by the same update rule of GLVQ for reducing the number of misclassification on training samples. In a test phase, a test sample is classified by LSC with trained prototypes. Experimental results on a handwritten digit show that the proposed learning rule outperforms other classifiers such as the original GLVQ algorithm.