Deep connectomics networks: Results from neural network architectures inspired from network neuroscience

Claims from the fields of network neuroscience and connectomics suggest that topological models of the brain involving complex networks are of particular use and interest. The field of deep neural networks has mostly left inspiration from these claims out. In this paper, we propose three architectures and use each of them to explore the intersection of network neuroscience and deep learning in an attempt to bridge the gap between the two fields. Using the teachings from network neuroscience and connectomics, we show improvements over the ResNet architecture, we show a possible connection between early training and the spectral properties of the network, and we show the trainability of a DNN based on the neuronal network of C.Elegans.