Restricted Convolutional Neural Networks

摘要

In this paper, a new type of convolutional neural network is proposed which is inspired by cellular automata research. This model is referred to as “restricted convolutional neural network” and its characteristic is that the feature maps are not fully connected, i.e. each feature map is only connected to a small neighborhood of previous feature maps. First this model is formally defined. Then it is used for image classification. Two layerwise pretraining methods have been proposed, and some structural variations have been analyzed. The model is tested on both MNIST and CIFAR-10 datasets. Results suggest that this model in some cases can outperform a convolutional neural network with similar architecture.