Deep Neural Networks with Cascaded Output Layers

This paper presented a novel design method for deep neural network aiming at promoting its generalization performance. First, the common methods for improving its generalization performance were introduced. Then, the deep neural networks with cascaded output layers was described. This deep neural networks have several output layers that constitute sequential classifiers, which combined with the adaboost algorithm, boost the generalization performance. Simultaneously, the proposed design method saves the computation consumption by sharing partial network structure among these classifiers. Finally, experiments were conducted to confirm the effectiveness and reliability of the method. © 2017, Editorial Department of Journal of Tongji University. All right reserved.