Online Visualization of Prototypes and Receptive Fields Produced by LVQ Algorithms

A new approach is proposed to visualize online the training of learning vector quantization algorithms. The prototypes and data samples associated to each receptive field are projected onto a two-dimensional map by using a non-linear transformation of the input space. The mapping finds a set of projection vectors by minimizing a cost function, which preserves the local topology of the input space. The proposed visualization is tested on two datasets: image segmentation and pipeline. The usefulness of the method is demonstrated by studying the behavior of Generalized LVQ, Supervised Neural Gas and Harmonic to Minimum LVQ algorithms on high-dimensional datasets.