Mapping, Implementing, and Programming Spiking Neural Networks

Abstract

Computer architectures inspired by biological neural networks are currently an area of growing interest, due to immense utility of these systems which is shown by their near-ubiquity within animals. An essential aspect of these systems is their ability to compute through the exchange of temporal events called ‘spikes.’ However, many aspects of biological computation remain unknown. To improve our ability to measure neural systems, we create an efficient implementation and statistical testing method to calculate an information-theory based metric, transfer entropy, on signals recorded from cultures of neurons. Taking inspiration from established knowledge regarding biological neurons, we investigate the impact which stochastic behavior has on the robustness of spiking networks when their synaptic weights are inaccurate. We find that a level of stochasticity can help improve this robustness. Lastly, we investigate methods of creating programs for spike-based computation through evolutionary optimization methods, and identify opportunities and challenges in this area.