ICA USING KERNEL CANONICAL CORRELATION ANALYSIS

Colin Fyfe and Pei Ling Lai

We derive a new method based on kernels for per- forming Canonical Correlation Analysis. We show that the method can be used to extract individual sinusoids from a linear mixture of sinusoids and that this is also possible when the number of mixtures is less than the number of signals, when there is a nonlinear mixture of the signals and when the mixture is time varying. In the last case, the nature of the time varying mixture matrix is revealed by some of the lower order canonical correlations when we use a nonlinear kernel.