Discriminative Clustering: Vector Quantization in Learning Metrics

Discriminative clustering uses auxiliary data to discover task-relevant characteristics of primary data. Asymptotically such clustering is equivalent to vector quantization in the primary data space, but with a new metric. It is implicitly assumed that changes in the primary data are important only to the extent they cause variation in the auxiliary data; the new metric is set up to measure the important changes. In this paper, optimization and regularization of discriminative clusters are discussed.