Next: Nonlinear decorrelation algorithms
Up: Algorithms for ICA
Previous: Preprocessing of the data
The pioneering work in [80] was
inspired by neural
networks. Their algorithm was based on canceling the
nonlinear crosscorrelations, see Section 4.3.3.
The nondiagonal terms of the matrix
are updated according
to

(33) 
where g_{1} and g_{2} are some odd nonlinear functions, and the y_{i}are computed at every iteration as
.
The
diagonal terms
are set to zero. The y_{i} then give, after
convergence, estimates of the independent components.
Unfortunately, the algorithm converges only under
rather severe restrictions (see [40]).
Aapo Hyvarinen
19990423