jdm5548@diamond.tamu.edu (06/22/91)
I've seen discussion here before about fancy normalization procedures (computing the pdf of each input vector and warping it to be Gaussian). Not to be fancy, what is the *simple* but yet preferred way to normalize for bp (my transfer function is [-1,1] sigmoidal, if it matters)? (a) scaling (dividing each element by the max element in the vector) (b) normalization (dividing each element by the length of the vector) (c) global normalization (dividing every element in every input vector by the maximum length of all input vectors in a training set) (d) something else I'm going with (b) for now, but I have a hunch that (c) might be better. --jdm5548@diamond.tamu.edu, jdm5548@tamagen.bitnet