Computation Complexity Reduction Based on Quick Leaky-Integrate-And-Fire Mechanism for SNNs

SNNs excel in processing event-based information with significantly higher energy efficiency, demonstrating great potential in resource-constrained edge computing systems. In this paper, we propose an efficient SNN architecture that uses quick leaky-integrate-and-fired ($q L I F$) mechanism. qLIF expedites the leakage of the membrane potentiation in the neuron, so the firing of spikes is less dependent on past spike-train events, instead, it is more relevant to the accumulated arrival of the current inputs. Based on this mechanism, the convolutional computation complexity can be greatly optimized, resulting in a dramatic acceleration of the convolutional operation and lower energy consumption, while maintaining a limited performance loss. A revised SNN based on qLIF is implemented in Verilog HDL and then synthesized using the 55 nm CMOS process library running on a 50 MHz clock. The energy consumption per inference on the MNIST dataset and the MIT-BIH dataset reaches $0.142 \mu \mathrm{~J} / 0.032 \mu \mathrm{~J}$, and the inference time is $0.150 \mathrm{~ms} / 0.053 \mathrm{~ms}$, respectively. Our proposed processor outperforms other state-of-the-art SNN processors in terms of latency and energy efficiency during inference computation.