SAPIENS: Towards Energy-efficient Deep Spiking Neural Networks via the Approximation of Power-expensive Neurons

We present an energy-saving technique for deep spiking neural networks (SNNs) converted from pre-trained analog neural networks (ANNs), named SAPIENS. Deep SNNs have many spiking operations, causing considerable switching power dissipation in the systems that process SNNs. Our study shows that in deep SNNs, only 20% of neurons account for more than 80% of SNN operations, called expensive neurons. Our SAPIENS technique identifies and approximates expensive neurons on-the-fly, improving energy consumption by 50–60% while having minimal impact on SNN accuracy.