Topographic property of backpropagation artificial neural network: From human functional connectivity network to artificial neural network

Abstract

The human brain network is one of the most complex data processing centers so far and shows several topographic features such as small worldness, modularity, and high efficiency. Inspired by the human brain network, the artificial neural network (ANN) method was proposed and has become an important engineering tool for pattern recognition. A recent study suggested the similarity between the human brain network and ANN. Would the ANN show topographic features that are similar to those of the human brain network? In the present study, we first constructed a simple backpropagation ANN (BP-ANN) with one hidden layer, which is one of the simplest, representative ANN, based on CIFAR10 and CIFAR100 datasets. Then, we explored the topographic features of a network with artificial neurons. We found that artificial neurons of the BP-ANN showed a high local efficiency and modularity similar to the human brain network. However, the global efficiency of artificial neurons of the BP-ANN was comparable with those of a random network. Our results showed that the best model (high accuracy and no overfitting) of a large-scale BP-ANN during epochs of training showed the lowest local and global efficiency. These results might imply that after epochs of training, neurons of the BP-ANN form local highly intra-connected functional modules but lack cooperation at the global level. The topography structure of the BP-ANN also shows the potential to become an early-stop marker during training.