Selective sampling for reliable neural signal approximation

ANN learning poses restrictions on the learning algorithm in combination with the structure of the training set. We analyse such a restriction as originating from the long term effect on learning of so-called cancelation examples. It is pointed out that cancelation effects may creep in unnoticed leading to non-reproducible and large learning times for real-life measurements. A selective sampling strategy is proposed to guarantee even for these cases a high-quality, stable learning as illustrated in the diagnosis of power generators