Sparse connectivity enables efficient information processing in cortex-like artificial neural networks

Neurons in cortical networks are very sparsely connected; even neurons whose axons and dendrites overlap are highly unlikely to form a synaptic connection. What is the relevance of such sparse connectivity for a network's function? Surprisingly, it has been shown that sparse connectivity impairs information processing in artificial neural networks (ANNs). Does this imply that sparse connectivity also impairs information processing in biological neural networks? Although ANNs were originally inspired by the brain, conventional ANNs differ substantially in their structural network architecture from cortical networks. To disentangle the relevance of these structural properties for information processing in networks, we systematically constructed ANNs constrained by interpretable features of cortical networks. We find that in large and recurrently connected networks, as are found in the cortex, sparse connectivity facilitates time- and data-efficient information processing. We explore the origins of these surprising findings and show that conventional dense ANNs distribute information across only a very small fraction of nodes, whereas sparse ANNs distribute information across more nodes. We show that sparsity is most critical in networks with fixed excitatory and inhibitory nodes, mirroring neuronal cell types in cortex. This constraint causes a large learning delay in densely connected networks which is eliminated by sparse connectivity. Taken together, our findings show that sparse connectivity enables efficient information processing given key constraints from cortical networks, setting the stage for further investigation into higher-order features of cortical connectivity.