ATW-SNN: low-accuracy-loss asymmetric ternary-weight spiking neural network with spike calibration strategy

Abstract The deployment of spiking neural networks (SNNs) on resource-constrained devices is mainly limited by memory footprint and energy consumption. Specifically, these limitations stem from the overhead incurred by full-precision storage and floating-point operations. To address these bottlenecks, we propose the asymmetric ternary weight SNN (ATW-SNN): by quantizing weights into asymmetric ternary values, ATW-SNN effectively reduces memory requirements and the number of floating-point operations. To mitigate the accuracy loss induced by quantization, we introduce a spike calibration strategy. This strategy leverages cosine distance to evaluate the discrepancy in spike distribution between ATW-SNN and the full-precision SNN, and conducts knowledge distillation based on this discrepancy. This process is designed to align the output of ATW-SNN with that of the full-precision SNN, thereby maintaining comparable performance. Moreover, we design an inference process for ATW-SNN on edge devices that changes the computing paradigm to utilize sparse weights and integer accumulation for enhanced energy efficiency. We conducted experiments on CIFAR-10, CIFAR-100, and DVS-Gesture, and the results show that the accuracy loss is less than 1%, while the weight memory is compressed by at least 14.91× and energy consumption is reduced by at least 5.95 ×. This work provides a feasible solution for the energy-efficient deployment of SNNs on edge devices.