A novel batch training algorithm for learning vector quantization networks using soft-labeled training data and prototypes

Learning vector quantization (LVQ) network is a prototype based classifier and known as a supervised neural network. Prototypes in LVQs represent the feature of the classes in data. In order to train and adjust network parameters in an offline manner, it is necessary to apply a huge amount of the training samples. In such cases, batch algorithms are efficient. So, this paper concentrates on the existing batch training algorithms and robustness of the LVQs using soft-labeled training data and prototypes. The paper assumes that training sets originally have hard labels where each sample is exclusively associated with a specific class. Hence at first, the samples are converted to soft labels using Keller method. Finally, a novel batch training algorithm is proposed that not only updates the network parameter but also adapts the neighborhood function during training phase. Simulation results show the performance of the proposed algorithm such as convergence rate and data classification accuracy.