Correlation Cascade and maximazing correlation

Hello Every body,

I have been doing an implementation of the cascade correlation as
published by S. Fahlman et al. and it works pretty good but I decided
to change somethings of the standard system.

So instead of maximazing the correlation of each NN with the last
error I minimize the square mean error of the output of the individual
NN with the error of that step. In principle thinks should work nearly
the same but I am having very hard problems to train the output unit
(output SLP). I need a lot of initializations and the results aren´t
so good as maximazing the correlation.

¿Does someone study this problem? ¿Why doesn´t work minimazing the
Mean Squared Error?

Thank you in advance,
Pedro Cabrera