PTQ4SNN: Membrane-Aware Post-Training Quantization for Spiking Neural Networks
There's a new preprint on arXiv (2608.07066v1) that introduces PTQ4SNN, a framework aimed at improving post-training quantization for spiking neural networks (SNNs). It tackles a major problem: when SNNs are deployed with low-bit precision, the membrane states often stay in floating-point format even after the weights are quantized. This situation arises because of differences in how channels distribute and their sensitivity near the firing threshold. PTQ4SNN uses a small calibration set and incorporates two key components: a channel-wise Unified Scale Bridge, which facilitates compatible conversions, and Mixed-Precision Bit Allocation, which tailors the bit precision (2/4/8 bits) based on firing activity and sensitivity, ultimately boosting the efficiency and accuracy of SNN deployment.
Key facts
- PTQ4SNN is a post-training quantization framework for spiking neural networks.
- It jointly quantizes weights and recurrent membrane states.
- The method uses a small calibration set.
- The Unified Scale Bridge constrains membrane scale as s_mem,c = s_w,c * 2^k_c.
- Mixed-Precision Bit Allocation assigns 2/4/8-bit precision to membrane channels.
- The paper is available on arXiv with ID 2608.07066v1.
- The approach addresses the challenge of quantizing membrane states in SNNs.
- The method aims to enable low-bit deployment of SNNs.
Entities
Institutions
- arXiv