| Neuromorphic engineers have achieved considerable success in devising silicon implementations of progressively more complex neural architectures. However, the effort required to design a successful neuromorphic system grows dramatically as the scope of these projects expands to encompass multiple neuromorphic subsystems. This design process could be eased by automating difficult design tasks.; In this thesis I introduce a novel technique for automatically rewiring connectivity between spiking neurons based on a model of activity-dependent axonal growth cone navigation during neural development, and illustrate its performance with a silicon implementation of a model growth cone population whose migration is driven and directed by patterned neural activity. I develop a stochastic model of silicon growth cone motion to explain and characterize population behavior, and discover that performance is limited by an optimality criterion whose existence is implied by the fundamental physicality of the system. |
