
Regression
and Neural Networks
169
balanced, gigantic values that squeeze the last bit of fit out of the
training set, but Eire worthless in the general population.
It is unfortunate that determination of a "good" cutoff on the
minimum weight allowed is fairly arbitrary and problem dependent.
Certainly we are always safe in discarding only zeros, as those
contribute nothing. That alone is a major improvement over the
traditional regression method, as it sacrifices nothing, but gains the
important advantage of eliminating all problems of singular matrices.
Another generally agreed-upon principle is that the cutoff must be set
relative to th