Abstract
We performed a massive evaluation of neural networks with architecturescorresponding to random graphs of various types. Apart from the classicalrandom graph families including random, scale-free and small world graphs, weintroduced a novel and flexible algorithm for directly generating randomdirected acyclic graphs (DAG) and studied a class of graphs derived fromfunctional resting state fMRI networks. A majority of the best performingnetworks were indeed in these new families. We also proposed a generalprocedure for turning a graph into a DAG necessary for a feed-forward neuralnetwork. We investigated various structural and numerical properties of thegraphs in relation to neural network test accuracy. Since none of the classicalnumerical graph invariants by itself seems to allow to single out the bestnetworks, we introduced new numerical characteristics that selected a set ofquasi-1-dimensional graphs, which were the majority among the best performingnetworks.