为多模态分类任务定制脉冲神经网络的衰减时间常数，提升能效与精度。
Task-Aware Tuning of Time Constants in Spiking Neural Networks for Multimodal Classification
- 根据任务类型动态调整脉冲神经元的衰减时间常数以优化性能。
- 静态与动态图像任务中，中等时间常数使准确率最高，权重分布更集中。
- 时序数据任务需更宽的时间常数范围以保留长期特征，适合硬件容错设计。
脉冲神经网络（SNNs）在可穿戴传感和时序分析等低功耗边缘计算领域具有潜力。关键神经参数——漏电时间常数（LTC）决定了漏电积分-放电（LIF）神经元对信息的时序整合能力，但其在不同数据模态下的影响仍不明确。本研究考察了在静态图像、动态图像和生物信号时序分类任务中，采用时序自适应前馈SNN时LTC的作用。实验表明，LTC显著影响推理准确率、突触权重分布和放电动力学。对于静态与动态图像，中等LTC获得更高准确率，且权重直方图紧凑居中，反映稳定的特征编码；在时序任务中，最优LTC增强时序特征保留，并导致更宽的权重稀疏性，具备对LTC变化的容忍度。结果表明，推理准确率在特定LTC范围内达到峰值，超出该范围则因过度整合或过快遗忘导致显著下降。放电率分析揭示了LTC、网络深度与能效间的强关联，强调平衡脉冲活动的重要性。这些发现表明，任务特异性LTC调优对高效脉冲编码与鲁棒学习至关重要。研究为硬件感知的SNN优化提供了实用指导，展示了如何设计神经元时间常数以匹配任务动态。本工作推动了面向实时分类任务的可扩展、超低功耗SNN在类脑计算中的部署。
原文摘要 · Abstract (English)
Spiking Neural Networks (SNNs) are promising candidates for low-power edge computing in domains such as wearable sensing and time-series analysis. A key neuronal parameter, the leaky time constant (LTC), governs temporal integration of information in Leaky Integrateand-Fire (LIF) neurons, yet its impact on feedforward SNN performance across different data modalities remains underexplored. This study investigates the role of LTC in a temporally adaptive feedforward SNN applied to static image, dynamic image, and biosignal time-series classification. Presented experiments demonstrate that LTCs critically affect inference accuracy, synaptic weight distributions, and firing dynamics. For static and dynamic images, intermediate LTCs yield higher accuracy and compact, centered weight histograms, reflecting stable feature encoding. In time-series tasks, optimal LTCs enhance temporal feature retention and result in broader weight sparsity, allowing for tolerance of LTC variations. The provided results show that inference accuracy peaks at specific LTC ranges, with significant degradation beyond this optimal band due to over-integration or excessive forgetting. Firing rate analysis reveals a strong interplay between LTC, network depth, and energy efficiency, underscoring the importance of balanced spiking activity. These findings reveal that task-specific LTC tuning is essential for efficient spike coding and robust learning. The results provide practical guidelines for hardware-aware SNN optimization and highlight how neuronal time constants can be designed to match task dynamics. This work contributes toward scalable, ultra-lowpower SNN deployment for real-time classification tasks in neuromorphic computing.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。