Structural Plasticity in Recurrent Neural Networks: Multi-Contact Formation and Homeostatic Network Rewiring

The brain consists of neurons connected by synapses which are highly plastic. This plasticity can appear by changing the strength (size) of synapses, but also by creating new and pruning existing connections. Here we analyze neocortical circuits that exhibit such structural plasticity, using numerical simulations of large-scale recurrent networks of integrate-and-fire neurons and mathematical modeling. First we explore a model of plasticity which is driven by precise spike correlations between pre- and post-synaptic neurons. We show that this model can account for the experimentally reported distributions of synaptic weights and numbers of synaptic contacts. In the second part we explore a structural plasticity model based on simple firing rate homeostasis, which has previously been shown to have interesting Hebbian properties. We developed a mean-field theory to analyze the ability of network to store memories in its connectivity, and its relation to the standard Hebbian paradigm with explicit correlation dependence. We show that mean-field theory does predict the formation of structure. Finally we speculate about an expansion of the mean-field theory to explain the decay of structure through diffusion.