An Efficient Hardware Implementation of Spiking Neural Network Using Approximate Izhikevich Neuron

Implementation of a highly complicated spiking neuron model on the hardware often requires a large amount of hardware resources. To deal with this problem, the mathematical accuracy of the neuron model needs to be compromised to allow for a more efficient implementation. In this paper, we utilize Approximate Computing to achieve the trade-offs between computational accuracy and hardware requirements. Instead of naively applying approximate adders and approximate multipliers to the Izhikevich model, this work explores different combinations of adders and multipliers in conjunction with an approximate aware approach. The evaluation results of STDP-trained Spiking Neural Networks with 6,400 neurons show that with a proper approximation technique, the highest accuracy of the approximate neuron model is $81.9 \%$, which suffers a $1.5 \%$ accuracy loss but reduces area cost and power consumption of the neuron by $40.3 \%$ and $24.2 \%$ respectively when compared to the accurate version of the Izhikevich neuron model.