Post-training quantization for efficient ANN-SNN conversion.

Abstract

Spiking Neural Networks (SNNs), inspired by the behavior of biological neurons, offer a promising direction for next-generation neural computing. Two primary methodologies have emerged for training deep SNNs: Direct Training, which optimizes SNNs using surrogate gradients, and ANN-to-SNN Conversion, which derives SNNs from Artificial Neural Networks (ANNs). In this work, we focus on the latter and investigate the conversion error that arises during the transformation. We provide a theoretical analysis showing that channel-wise thresholds are more effective than traditional layer-wise thresholds in mitigating this error. To achieve this efficiently, we leverage post-training quantization (PTQ), which enables calibration using only a small dataset without requiring retraining. Compared to conventional direct training and ANN-to-SNN conversion methods, our approach significantly reduces training time while improving accuracy on both static image and neuromorphic datasets.