Selection of Synaptic Connections by Wiring Plasticity for Robust Learning by Synaptic Weight Plasticity

In the cortex of adult rodents, 5-15% of spines are created and eliminated every day, and the resultant synaptic connection structure in local neural circuits are highly nonrandom, suggesting crucial roles of synaptic rewiring plasticity in learning. However, it has not been known what the functional advantage is of having synaptic wiring plasticity on top of rich synaptic weight plasticity mechanisms and how stochastic creation and elimination of spines achieves that. To answer these questions, we constructed a simple computational model of an inference task and studied functions of synaptic rewiring by analytical and numerical methods. We found that a Hebbian-type wiring plasticity selectively eliminates noninformative connections to enhance learning by synaptic weight plasticity. Moreover, wiring plasticity enables rapid learning through weight plasticity by generating a noise-resilient connection structure in rapidly changing environments.