Practical Neural Network Recipies in C++

Preparing

Input Data

267

algorithms impose fixed limits on their weights to prevent getting

stuck out in the far forty. These algorithms will simply not be able to

learn such extreme data. We can make the network's life a lot easier

by giving it data scaled in such a way that all weights can remain in

small, predictable ranges.

This becomes even more important when the data is being

learned by output neurons. Most training algorithms minimize the

total error of all outputs. If the output variables are unequally scaled,

those with larger variabilities will be favored, as they will dominate

the error sum. This can have profound consequences an ...