next up previous
Next: LOCAL CONDITIONED VARIANCE MAXIMIZATION Up: OBJECTIVE FUNCTIONS FOR THE Previous: REMOVING THE GLOBAL INVERTIBILITY

A DISADVANTAGE OF THE ABOVE METHODS

Note that a factorial code causes non-maximal $V$ and therefore non-maximal $T$ for all methods with $\alpha > 0$ except for rare cases (such as if there are $2^n$ equally probable different input patterns). This means that with a given problem there is some need for parameter tuning of the relative weighting factors, due to the possibility that the various constraints may not be satisfiable simultaneously (see the footnote of section 4.4). The method in the next section avoids this necessity for parameter tuning by replacing the term for variance maximization by a predictor-based term for conditioned variance maximization.



Juergen Schmidhuber 2003-02-13


Back to Independent Component Analysis page.