Abstract
To understand how intelligence is generated based on neural connections using neural network models, in addition to train Spiking Neural Networks (SNNs) for accomplishing a given cognitive task, another important topic is to investigate how SNNs generate meaningful activities in response to external stimuli. Previous studies have proposed that a low rank connectivity matrix in Recurrent Neural Networks (RNNs) can provide the key mechanism of computation ability of connected neurons. However, these studies are unconvincing when applied to real neurons because the neuron models used in these studies were overly simplistic and lacked adequate biological plausibility such as the definition of reversal potential and excitatory-inhibitory interaction. Addressing this lack, our research develops more biologically plausible SNNs to reveal the origin of computation ability in neuron populations. Here we use Voltage Dependent Theta Neuron Model to construct SNNs with low rank connectivity. In addition to connectivity parameters, we also investigate the effect of those biological parameters like different synaptic decay times and inhibitory connections on the performance of SNNs. Consequently, the findings from this research have greater applicability to real neural connections.
