Te-Won Lee, (Joint work with Taesu Kim and Intae Lee)
Institute for Neural Computation, University of California, San Diego
We present a new approach to independent component analysis by extending the formulation of a single variable source signal to a vector source signal. This approach is therefore termed independent vector analysis or IVA. The independence cost function is a modified form of the Kullback-Leibler divergence function in which the difference between the joint probability of multidimensional sources and the product of marginal multidimensional probabilities is measured. The model allows independence between multidimensional sources represented as vector sources and dependence between the source signals that form the vector representation. A multidimensional density model can for example capture variance dependencies among the source signals in the vector. The learning algorithm resembles the well known infomax algorithm and differs only in the form of the score function which depends on multiple source estimates instead of a single source estimate. There are several applications of this formulation. For the separation of acoustic sources, the algorithm mitigates the permutation problem, i.e. most ICA algorithms applied in to the frequency domain mixture data suffer from the unknown permutation of the output signals. Although there are several engineering solutions to fix this problem after the ICA stage, the proposed method provides a natural solution to the problem by capturing the inherent dependencies of the acoustic signals. It therefore avoids the permutation problem and allows the separation of sources in very challenging environments for many sound sources.