Biologically inspired artificial neural network for higher performance and robustness.

While biological neural networks exhibit much richer organization as compared to, $e.g.$, multilayer perceptrons (MLPs -- made of chain of bicliques, never observed in biological organisms), this higher and more efficient representational capability is not exploited in common network design strategies. In order to exploit the topological features selected by the evolutionary pressure, we propose the Elegant Neural Network (ENN), a feedforward neural network which shares its underlying undirected graph, at the microscopic scale, with the connectome of the $C. elegans$ nematode. Our aim is twofold: to contribute to the advancement of neural architecture search and to provide connectionists with a valuable simulation tool. We train both an ENN and a MLP, under the same number of neurons and parameters, to perform discrimination tasks with multidimensional input, and test the robustness of each network to node removal; ENNs either match or surpass the equivalent MLPs in terms of accuracy. Results indicate a complex interplay between topology, performance and robustness, and point towards novel, bio-inspired strategies in generating optimal neural architectures.