Both the multiplicative and z-axis normalization types allow a group of
outputs to be adjusted so that the "vector length" is 1. Both go about it in
different ways. Certain types of neural networks require a vector length of
Z-Axis normalization is usually a better choice than multiplicative. However,
multiplicative can perform better than Z-Axis when all of the values are near
zero most of the time. This can cause the "synthetic value" that z-axis uses
to dominate and skew the answer.
Z-Axis gets its name from 3D computer graphics, where there is a Z-Axis
extending from the plane created by the X and Y axes. It has nothing to do
with z-scores or the z-transform of signal theory.
To implement Z-Axis normalization a scaling factor must be created to
multiply each of the inputs against. Additionally, a synthetic field must be
added. It is very important that this synthetic field be added to any z-axis
group that you might use. The synthetic field is represented by the