Membrane-Aware Quantization Boosts Spiking Neural Network Efficiency

Spiking Neural Networks (SNNs) are promising for efficient, event-driven AI, but their deployment on low-bit hardware has been hindered by the need to keep recurrent membrane states in floating-point precision. Quantizing these states is complex due to varying distributions and the risk of altering spike decisions. Researchers have developed PTQ4SNN, a new post-training quantization method designed specifically for SNNs.PTQ4SNN addresses these challenges by introducing two key components. First, a Unified Scale Bridge adapts membrane scales to their unique distributions while ensuring compatibility with weight quantization. Second, a Mixed-Precision Bit Allocation strategy assigns optimal bit precision (2, 4, or 8-bit) to membrane channels based on their activity and sensitivity, all within a specified bit budget. This framework works with various SNN architectures, including convolutional SNNs and spike-driven Transformers, without requiring model retraining.Experimental results demonstrate that PTQ4SNN effectively preserves model accuracy, even with 4-bit weight quantization and approximately 4-bit membrane precision, across diverse tasks like static and event-based classification and semantic segmentation. This advancement significantly improves the feasibility of deploying SNNs on resource-constrained devices.