The effect of biologically-inspired mechanisms in spiking neural networks for neuromorphic implementation

An open question in neuromorphic computing is what neuron and synapse models should be used and what level of biological detail should be included in those models. For neuromorphic systems, the complexity level of the neuron and synapse model have a corresponding effect on the complexity of the hardware, and thus potentially affect the scale of network that can be feasibly implemented as well as the device's efficiency. In this work, we seek to determine what effect the inclusion of two biological mechanisms have on performance. We start with a basic spiking neural network model that currently has multiple hardware implementations and examine the effect of including two additional biologically-inspired mechanisms (simple potentiation/depression weight-change mechanisms on the synapses and leak on the neurons). We compare the performance of networks with and without these mechanisms on three simple applications. We found that, in general, these two mechanisms had relatively little effect on training and generalization performance, indicating that they may be able to be omitted in future neuromorphic implementations.