Adaptive estimation of parameters using partial information of desired outputs

Abstract

A general framework for forming an adaptive algorithm for problems where only partial information about the desired output is available, is proposed. Based on preliminary analysis it can be shown that this framework can be used to efficiently choose deep, narrow minima when there are many local minima. For problems like separation of instantaneous mixtures (independent component analysis, ICA) and separation of convolutive mixtures when cast in the proposed framework is shown to give the same efficient algorithms as those available in the literature.