Learning Hierarchical Representations with Spike-and-Slab Inception Network

Recently, deep convolutional neural networks (CNN) with inception modules have attracted much attention due to their excellent performances on diverse domains. Nevertheless, the basic CNN can only capture a univariate feature, which is essentially linear. It leads to a weak ability in feature expression, further resulting in insufficient feature mining. In view of this issue, researchers incessantly deepened the network, bringing parameter redundancy and model over-fitting. Hence, whether we can employ this efficient deep neural network architecture to improve CNN and enhance the capacity of image recognition task still remains unknown. In this paper, we introduce spike-and-slab units to the modified inception module, enabling our model to capture dual latent variables and the average and covariance information. This operation further enhances the robustness of our model to variations of image intensity without increasing the model parameters. The results of several tasks demonstrated that dual variable operations can be well-integrated into inception modules, and excellent results have been achieved.