Vector quantization using survival Cauchy-Schwartz divergence

Lei Guo, Hua Qu, Ji-hong Zhao, Badong Chen

Abstract: Vector quantization (VQ) is a data compression method in machine learning and data mining field by representing a larger data set with a smaller number of vectors in a possible way. Several vector quantization algorithms have been proposed in recent years. Different from the classic vector quantization algorithms such as LBG and K-means, the algorithms based on information theoretic learning try to make the code book distribution similar to the original data distribution. The computational complexity of such algorithms is however very high. In this paper, a novel vector quantization algorithm is proposed, which is based on a parameter-free information theoretic quantity, namely the survival Cauchy-Schwartz divergence (SCSD). Minimizing the SCSD between code book and the original data set yields a code book that has similar distribution with the data set. The computational cost of the new algorithm is relatively small due to the computational simplicity of the survival information potential (SIP). Simulation results show that the proposed algorithm works well.

External IDs:dblp:conf/mlsp/GuoQZC14