We present a novel generalized learning vector quantization (LVQ) framework called the harmonic to minimum generalized LVQ algorithm (H2M-GLVQ). Through incorporating the distance measure transition procedure from harmonic average distance to minimum distance, the H2M-GLVQ cost function is gradually changing from the soft model to the hard model. Our proposed method, at the early training stage, can effectively tackle the initialization sensitivity problem associated with the original generalized LVQ algorithm while the convergence of the algorithm can be ensured by the hard model in the later training stage. Experimental results have shown the superior performance of the H2M-GLVQ algorithm over the generalized LVQ and one of its variants on some artificial multi-modal datasets
Masaki Nakagawa,et al.
Evaluation of prototype learning algorithms for nearest-neighbor classifier in application to handwritten character recognition
Biing-Hwang Juang,et al.
Discriminative learning for minimum error classification [pattern recognition]
IEEE Trans. Signal Process..
Atsushi Sato,et al.
Generalized Learning Vector Quantization
Discriminative Dimensionality Reduction Based on Generalized LVQ
Barbara Hammer,et al.
Supervised Neural Gas for Learning Vector Quantization
An analysis of initial state dependence in generalized LVQ
Katsuhiko Takahashi,et al.
A class-modular GLVQ ensemble with outlier learning for handwritten digit recognition
Seventh International Conference on Document Analysis and Recognition, 2003. Proceedings..