Exploiting prior knowledge in network optimization: an illustration from medical prognosis

摘要: Feed-forward layered networks trained on a pattern classification task in which the number of training patterns each class is non-uniform, exhibit strong bias towards those classes with largest membership. This an unfortunate property when relative importance smaller membership much greater than that many patterns. In addition, there are tasks where different penalties associated misclassifying belonging to one as another class. Generally, it not known how compensate for such effects network training. paper discusses analytical regularization scheme whereby prior expectations occurring generalization data and misclassification costs may be incorporated into phase, thus compensating uneven unfair distributions set. The proposed feature extractio...