DiffusionNet: Establish convolutional networks with nitric oxide diffusion model

Abstract

Skip connections are used in DenseNets recently and have significantly improved network performance. In this paper, we compare skip connections with the diffusion process of endogenous Nitric Oxide (NO) between neurons, and propose DiffusionNets by replacing skip connections with NO diffusion model. Each layer is considered as a point spreading signal to space as well as receiving signal from space. The whole network transmits information with a diffusing way. DiffusionNets have several advantages: (1) generate more discriminative features. (2) more similar to neural information transmission. (3) higher classification accuracy. DiffusionNets were evaluated on CIFAR10 and CIFAR100 and outperform the original DenseNets.