Deep extractive networks for supervised learning

Abstract This paper introduces a new type of deep learning method named deep extractive networks (DEN) for supervised learning, which inherits the feature extraction ability of DBN and fuzzy representation ability of quantum neural networks (QNN). At first, we propose the architecture of DEN, which consists of quantum neuron and sigmoid neuron, can divide the samples of different classes into different areas in new Euclidean space. The parameter space of the deep architecture is initialized by greedy layer-wise unsupervised learning, and the parameter space of quantum representation is initialized with zero. Then, the parameter space of the deep architecture and quantum representation are refined by supervised learning based on the gradient-descent procedure. An exponential loss function is used in this paper to guide the supervised learning procedure. Experiments conducted on standard datasets show that DEN outperforms existing feedforward neural networks and neuro-fuzzy classifiers.