On the Improvement of a Linear Separation by Extending the Adaptive Process with a Stricter Criterion

A process is described for continuing adaptation, after hyperplanes that separate pattern classes in pattern space have been found, in order to increase the distance between sample patterns and hyperplanes. It is then shown that the original adaptive process is a special case of a more general adaptive process in which the distance of a mis-assigned training pattern from the border of its proper category region is regarded as negative. This device permits the concept of nearly linearly separable classes.