Evaluation of Neuron Parameters on the Performance of Spiking Neural Networks and Neuromorphic Hardware

There are a large variety of different neuron implementations that have been used in past spiking neural networks and neuromorphic hardware implementations. However, it has not been clear what the impact of neuron parameters and characteristics are on application and hardware performance. In this work, we investigate the impact of different neuronal features, including leak, and absolute and relative refractory periods, as well as their associated parameters, on application performance. We investigate the performance across a variety of real-world applications, including static and temporal classification tasks and control tasks, as well as two different algorithmic approaches, evolutionary algorithms, and reservoir computing. We analyze the impact of these neuronal features in terms of performance on the application, as well as energy usage.