Direct optimization of margins improves generalization in combined classifiers

00 1Cumulative training margin distributionsfor AdaBoost versusour "Direct Optimization OfMargins" (DOOM) algorithm.The dark curve is AdaBoost, thelight curve is DOOM. DOOMsacrifices significant training errorfor improved test error (horizontalmarks on margin= 0 line).1 IntroductionMany learning algorithms for pattern classification minimize some cost function ofthe training data, with the aim of minimizing error (the probability of misclassifyingan example). One example of such ...