跳转至

PTQ4SNN：面向脉冲神经网络的膜电位感知后训练量化技术

文章背景与核心概要

脉冲神经网络（SNNs）凭借其高效能、稀疏化和事件驱动的计算特性备受关注，然而，由于循环膜电位状态（recurrent membrane states）通常仍以浮点格式保留，其低比特部署长期以来一直面临瓶颈。本文介绍了一种名为 PTQ4SNN 的创新膜电位感知后训练量化框架，该框架仅需使用少量的校准数据集，即可在无需重新训练主干网络的情况下，实现对权重和膜电位的联合量化。

PTQ4SNN 通过引入统一尺度桥接（Unified Scale Bridge）机制，适应膜电位分布的同时实现位移兼容的尺度转换；并通过混合精度位分配（Mixed-Precision Bit Allocation）策略，根据发放活动和量化敏感度动态分配位宽。该研究在静态、基于事件的分类以及语义分割任务上展现出了卓越的性能，有效在 W4 量化及约 4 比特膜电位精度下保持了模型准确率。

摘要 (Summary)

Spiking Neural Networks (SNNs) offer energy-efficient, sparse, and event-driven computation, but achieving low-bit deployment has traditionally been bottlenecked by recurrent membrane states remaining in floating-point format. PTQ4SNN introduces a novel membrane-aware post-training quantization framework that jointly quantizes weights and membrane states using only a small calibration set, without requiring backbone retraining.

Spiking Neural Networks (SNNs) offer energy-efficient, sparse, and event-driven computation, but achieving low-bit deployment has traditionally been bottlenecked by recurrent membrane states remaining in floating-point format. PTQ4SNN introduces a novel membrane-aware post-training quantization framework that jointly quantizes weights and membrane states using only a small calibration set, without requiring backbone retraining.

论文元数据 (Paper Metadata)

  • Title: PTQ4SNN: Membrane-Aware Post-Training Quantization for Spiking Neural Networks
  • Authors: Hui Xie, Tong Shi, Haotong Qin, Aishan Liu, Xiaode Liu, Jinyang Guo
  • Submitted: August 7, 2026
  • Primary Subject: Artificial Intelligence (cs.AI)
  • arXiv Identifier: arXiv:2608.07066
  • DOI: 10.48550/arXiv.2608.07066
  • Title: PTQ4SNN: Membrane-Aware Post-Training Quantization for Spiking Neural Networks
  • Authors: Hui Xie, Tong Shi, Haotong Qin, Aishan Liu, Xiaode Liu, Jinyang Guo
  • Submitted: August 7, 2026
  • Primary Subject: Artificial Intelligence (cs.AI)
  • arXiv Identifier: arXiv:2608.07066
  • DOI: 10.48550/arXiv.2608.07066

正文摘要 (Abstract)

脉冲神经网络（SNN）支持稀疏且事件驱动的计算，但其低比特部署一直未能完全实现，因为即使在权重被量化后，循环膜电位状态通常仍保留为浮点格式。对这些状态进行量化具有挑战性，因为它们的分布在各个通道之间以及与前置权重相比存在差异，并且在发放阈值附近的微小扰动可能会改变脉冲决策并随时间累积。

Spiking neural networks (SNNs) enable sparse and event-driven computation, but their low-bit deployment remains incomplete because recurrent membrane states are commonly retained in floating point even after weight quantization. Quantizing these states is challenging because their distributions differ across channels and from the preceding weights, while small perturbations near the firing threshold may alter spike decisions and accumulate over time.

我们提出了 PTQ4SNN，这是一个膜电位感知的后训练量化框架，它仅需使用一小批校准集，即可联合量化权重和循环膜电位状态。 1. 统一尺度桥接（Unified Scale Bridge）： 通道级尺度约束（\(s_{mem,c} = s_{w,c} \times 2^{k_c}\)），它既能适应膜电位分布，又能实现支持移位操作（shift-compatible）的尺度转换。 2. 混合精度位分配（Mixed-Precision Bit Allocation）： 在平均位宽预算下，根据发放活动和量化敏感度，为膜电位通道动态分配 2/4/8 比特精度。

We propose PTQ4SNN, a membrane-aware post-training quantization framework that jointly quantizes weights and recurrent membrane states using only a small calibration set. 1. Unified Scale Bridge: A channel-wise scale constraint (\(s_{mem,c} = s_{w,c} \times 2^{k_c}\)) that adapts to membrane distributions while enabling shift-compatible scale conversion. 2. Mixed-Precision Bit Allocation: Dynamically assigns 2/4/8-bit precision to membrane channels according to firing activity and quantization sensitivity under an average-bit budget.

该框架基于可重用的投影-LIF（Projection-LIF）对运作，并支持卷积 SNN 和脉冲驱动 Transformer，无需重新训练主干网络。在静态与基于事件的分类以及语义 segmentation（语义分割）任务上的实验表明，PTQ4SNN 在 W4 量化和约 4 比特膜电位精度下能有效保持模型准确率。

The framework operates on reusable projection-LIF pairs and supports both convolutional SNNs and spike-driven Transformers without backbone retraining. Experiments on static and event-based classification and semantic segmentation show that PTQ4SNN effectively preserves model accuracy under W4 quantization and approximately 4-bit membrane precision.