从无线脉冲神经网络到脉冲神经 P 系统：一种基于低能耗规则的转换
From Wireless SNNs to SN P Systems: A Low-Energy Rule-Based Conversion
浏览论文内容
中文总结 AI 辅助
本文针对分布式无线脉冲神经网络存在的问题，提出将其转换为脉冲神经 P 系统的方法，通过提取隐藏层脉冲活动规则实现。实验表明该系统能耗低且能保留约 84%准确率，输出层连接性降低，可作为轻量级、可解释替代物用于神经形态边缘部署。
中文摘要 AI 辅助
分布式无线脉冲神经网络（DWSNNs）是资源受限环境（如无线传感器网络）中节能边缘推理的一种有前景的范例。然而，存在内部决策过程不透明和剩余能量占用仍是超低功耗部署限制因素这两个局限。本文提出一种系统方法，通过从隐藏层脉冲活动中提取符号发放规则，将训练好的 DWSNN 转换为等效的脉冲神经 P（SN P）系统，这是一种受生物启发、基于膜计算的基于规则的计算模型。所得 SN P 系统能提供直接、人类可读的决策解释，且能耗比其母体 SNN 少三个数量级。在使用相位编码和泄漏积分发放（LIF）神经元的两层全连接 SNN 上对神经形态 MNIST（N-MNIST）数据集进行实验表明，SN P 系统保留了约 84%的原始分类准确率（73.77%对 87.68%），同时输出层连接性从 1000 降至 120 个特定类连接。这种复杂度降低由与每个类相关隐藏神经元数量有关的参数控制，可根据计算复杂度降低和输出准确率之间的权衡来选择。这些结果将 SN P 系统定位为神经形态边缘部署中训练好的分布式无线 SNN 的轻量级、可解释替代物。
英文摘要
Distributed wireless spiking neural networks (DWSNNs) are a promising paradigm for energy-efficient edge inference in resource-constrained environments such as wireless sensor networks (WSNs). Yet, two limitations persist: their internal decision process is opaque, and their residual energy footprint remains a limiting factor for ultra-low-power deployments. This paper proposes a systematic methodology to convert a trained DWSNN into an equivalent Spiking Neural P (SN P) system, a biologically-inspired, rule-based computational model drawn from membrane computing, by extracting symbolic firing rules from the hidden-layer spike activity. The resulting SN P system provides direct, human-readable decision explanations while consuming three orders of magnitude less energy than its parent SNN. Experiments on the Neuromorphic MNIST (N-MNIST) dataset with a two-layer fully connected SNN using phase encoding and Leaky Integrate-and-Fire (LIF) neurons show that the SN P system retains approximately 84% of the original classification accuracy (73.77% vs. 87.68%) while the output layer connectivity decreases from 1000 to 120 class-specific connections. This complexity reduction is governed by a parameter related to the number of relevant hidden neurons per class that can be chosen according to a trade-off between computational complexity reduction and output accuracy. These results position SN P systems as lightweight, interpretable surrogates for trained distributed wireless SNNs in neuromorphic edge deployments.
