Quantitative evaluation of dependence among outputs in ECOC classifiers using mutual information based measures

In previous works, it has been experimentally shown that the implementation of Error Correcting Output Coding (ECOC) classification methods with an ensemble of parallel and independent non linear dichotomizers (ECOC PND) outperforms the implementation with a single monolithic multi layer perceptron (ECOC MLP). The low dependence of the errors on different codeword bits was qualitatively indicated as one of the main factors affecting this result. In this paper, we quantitatively evaluate the dependence of output errors in ECOC learning machines using mutual information based measures, and we study the relation between dependence of output errors and classification performances.