PUB - Publikationen an der Universität Bielefeld

The Growing Neural Gas (GNG) algorithm is able to perform continuous vector quantization for an unknown distribution while preserving the topological structure of the input space. This makes the GNG attractive for online learning of visual codebooks. However, mapping an input vector to a reference vector is quite expensive and requires an iteration through the entire codebook. We propose a hierarchical extension of the Growing Neural Gas algorithm for online one-shot learning of visual vocabularies. The method intrinsically supports mapping input vectors to codewords in sub-linear time. Further, our extension avoids overfitting and locally keeps track of the topology of the input space. The algorithm is evaluated on both, low dimensional simulated data and high dimensional real world data.