A Framework for Metric Learning and Embedding with Topology Learning Neural Networks

A framework for metric learning and embedding with topology learning neural networks is proposed. To stress the problems of low efficiency in both time and space in conventional embedding methods such as Multi-Dimensional Scaling and Isomap, we take the advantage of incremental training and vector quantization abilities of topology learning neural networks such as Growing Neural Gas and Self-Organizing Incremental Neural Networks to construct a representation of the data. Then the embeddings are approximated with the graph similarities of the neurons instead of pairwise similarities of input data. In an experiment the proposed metric learning is used in combine of Support Vector Machine to solve a semi-supervised learning (SSL) problem. The results show that our proposed method increased classification accuracy in the SSL experiment.