DBNet: A New Generalized Structure Efficient for Classification
Yonghao Dang, Fuxing Yang, Baiquan Su, Jianqin Yin, Jun Liu · 2019
In this paper, we propose a new deep learning structure named deep-broad network (DBNet) that is efficient for classification task. By modifying the decision-making mechanism of the deep structure, the proposed method can improve the testing efficiency while maintaining the testing accuracy. Specifically, the modified convolutional neural networks (CNNs) are first pre-trained and used to extract high-quality features. And then the dimension of extracted features is reduced by linear mapping. Finally, the broad learning system (BLS) uses processed features to make decisions. Compared with the previous deep structure, the efficiency of the proposed model is improved. Compared with the BLS, the DBNet has better performance. The proposed model is evaluated by using the CIFAR-10, CIFAR-100 and MNIST datasets. And experimental results show that the DBNet is effective and efficient. Code and models are available at https://github.com/YHDang/DBNet.