Learning Parsimonious Deep Feed-forward Networks

Convolutional neural networks and recurrent neural networks are designed with network structures well suited to the nature of spacial and sequential data respectively. However, the structure of standard feed-forward neural networks (FNNs) is simply a stack of fully connected layers, regardless of the feature correlations in data. In addition, the number of layers and the number of neurons are manually tuned on validation data, which is time-consuming and may lead to suboptimal networks. In this paper, we propose an unsupervised structure learning method for learning parsimonious deep FNNs. Our method determines the number of layers, the number of neurons at each layer, and the sparse connectivity between adjacent layers automatically from data. The resulting models are called Backbone-Skippath Neural Networks (BSNNs). Experiments on 17 tasks show that, in comparison with FNNs, BSNNs can achieve better or comparable classification performance with much fewer parameters. The interpretability of BSNNs is also shown to be better than that of FNNs.