Who is afraid of the BIG bad ANN?

Ten-year experience with training large-scale ANN models with the PCA-CG algorithm that generates a non-random initial connection weight set is presented. The suggested small number of hidden neurons and the automatic identification and removal of the less relevant inputs increases the robustness of these models. Examples of ANN modeling of "artificial nose" sensor array, TV program rating and e-mail letter importance classification demonstrate the algorithm efficiency.