Convolutional networks with cross-layer neurons for image recognition

Very deep convolutional networks have recently achieved a series of breakthroughs on several challenging tasks such as the ImageNet or COCO competitions. However, it is difficult to train such deep neural networks. In this paper, we present a novel structure called cross-layer neurons architecture, which has the capability to train effective deeper neural networks. It utilizes cross-layer neurons to synthesize the information (features) learned from all the lower-level layers and send them to the higher-level layers through the cross-layer. Based on this novel architecture, we propose a new deep neural model termed Cross-Layer Neurons Networks (CLNN). It is shown that CLNN can relieve the problem of vanishing gradient. It is also shown that CLNN has the capability of improving the convergence rate of classification. Comparative experiments on several benchmark datasets (MNIST, CIFAR-10, CIFAR-100, SVHN and STL-10) clearly demonstrate that our proposed model is suitable for training deeper networks and can effectively improve the performance by utilizing cross-layer neurons.