ICA: A Critical Review of Three Prominent Approaches
Sumit Basu · 2000
We present the problem of IndependentComponents Analysis (ICA) and review three major approaches as described by Comon [9],Amari et al. [2], and Bell and Sejnowski [5]. We cast all three in a common notational framework, point out their strengths and weaknesses, and showhow they are related to eachother. We then go on to suggest several empirical studies to further investigate the numerical behavior of the algorithms. Finally, we presentanumber of novel extensions/applications of ICA. The most interesting of these is our notion of "conditionally independent components analysis," in whichwe propose factoring the conditional density p(xjy) to greatly reduce the amount of data needed to accurately model a scaled version of p(yjx).