Deep Error Correcting Output Codes

In this paper, we present a novel deep ensemble learning algorithm, called deep error correcting output codes (DeepECOC), which adapts traditional ECOC models to deep architectures. DeepECOC are composed of multiple layers of the ECOC module that combines multiple binary classifiers for feature learning. Unlike previous deep networks, such as deep autoencoders, supervisory information can be used during the greedy layer-wise pre-training of DeepECOC. We have conducted extensive experiments to compare DeepECOC with traditional ECOC, feature learning and deep learning algorithms on several benchmark data sets. The results not only show the advantages of DeepECOC over previous related approaches, but also demonstrate the effectiveness of supervised greedy layer-wise pre-training in contrast to unsupervised counterpart.