Less is different: why sparse networks with inhibition differ from complete graphs

Abstract: In neuronal systems, inhibition contributes to stabilizing dynamics and regulating pattern formation. Through developing mean field theories of neuronal models, using complete graph networks, inhibition is commonly viewed as one ``control parameter'' of the system, promoting an absorbing phase transition. Here, we show that for low connectivity sparse networks, inhibition weight is not a control parameter of the transition. We present analytical and simulation results using generic stochastic integrate-and-fire neurons that, under specific restrictions, become other simpler stochastic neuron models common in literature, which allow us to show that our results are valid for those models as well. We also give a simple explanation about why the inhibition role depends on topology, even when the topology has a dimensionality greater than the critical one. The absorbing transition independence of the inhibitory weight may be an important feature of a sparse network, as it will allow the network to maintain a near-critical regime, self-tuning average excitation, but at the same time, have the freedom to adjust inhibitory weights for computation, learning, and memory, exploiting the benefits of criticality.