Product approximation by minimizing the upper bound of Bayes error rate for Bayesian combination of classifiers

In combining multiple classifiers using a Bayesian formalism, a high dimensional probability distribution is composed of a class and decisions of classifiers. In order to do product approximation of the probability distribution, the upper bound of Bayes error rate, bounded by the conditional entropy of a class and decisions, should be minimized. A second-order dependency-based product approximation is proposed in this paper by considering the second-order dependency between the class and decisions. The proposed method is evaluated by combining the classifiers recognizing unconstrained handwritten numerals.