摘要
脉冲神经网络(Spiking Neural Networks,SNNs)是一种受大脑启发的计算模型,在处理具有时间和空间维度的数据时具有巨大潜力。研究表明,突触可塑性延迟能够提高SNNs在语音识别等时序任务的性能。然而,时序数据通常具有稀疏性和不均匀性,导致不同时间步中的信息分布存在显著差异。同时,信号中噪声的动态变化进一步加剧了这一问题,使得现有的可塑性延迟方法难以有效捕获关键时间信息,从而限制了模型性能的提升。在这项工作中,我们为可塑性延迟引入了时间注意力机制,通过动态评估各时间步的重要性,进而优化SNNs的性能。具体而言,我们使用扩展卷积来学习延迟和权重。随后,我们在时间维度上利用时间注意力机制动态量化不同时间步的重要性,从而增强模型对关键时间信息的捕获能力。实验结果表明,所提方法在Spiking Heidelberg Digits、Spiking Speech Command和Google Speech Commands数据集上分别取得了96.21%、80.87%和95.60%的精度,相较于传统的可塑性延迟方法有了显著的改进。这一结果验证了我们的方法在处理语音识别任务时的高效性,并为SNNs在处理复杂时序数据中的进一步应用奠定了坚实基础。
As a brain-inspired computational model,Spiking Neural Networks(SNNs)hold great promise for processing data with temporal and spatial dimensions.Research has shown that synaptic plasticity delays can improve the performance of SNNs in temporal tasks such as speech recognition.However,time-series data are usually sparse and inhomogeneous across time steps.This disparity in information distribution,coupled with noise dynamics in the signals,poses challenges for existing plasticity delay methods,which struggle to capture critical temporal information effectively,thereby limiting model performance.In this work,we apply a temporalwise attention mechanism to plasticity delays,enabling SNNs to evaluate the importance of each time step and optimize performance.Specifically,we use dilated convolution to learn the delay parameters and weights,followed by a temporal attention mechanism that dynamically quantifies the importance of different time steps,thereby enhancing the ability to capture critical temporal information of SNNs.Experimental results show that the proposed method achieves 96.21%,80.87%and 95.60%accuracy on the Spiking Heidelberg Digits,Spiking Speech Command and Google Speech Commands datasets,respectively,which is a significant improvement over traditional plasticity delay methods.These results highlight the potential of our approach for handling speech recognition tasks and establish a strong foundation for further applications of SNNs in dealing with complex time-series data.
作者
张航铭
白千一
邓智超
Alexander Sboev
于强
ZHANG Hang-Ming;BAI Qian-Yi;DENG Zhi-Chao;Alexander Sboev;YU Qiang(College of Intelligence and Computing,Tianjin University,Tianjin 300072;College of Computer and Information Engineering,Tianjin Normal University,Tianjin 300387;National Research Center Kurchatov Institute,Moscow 123182,Russia)
出处
《计算机学报》
北大核心
2025年第8期1870-1884,共15页
Chinese Journal of Computers
基金
国家自然科学基金(No.92370103,62176179)
小米公益基金会资助.
关键词
类脑计算
脉冲神经网络
突触可塑性延迟
时间注意力
神经形态计算
brain-like computing
spiking neural networks
synaptic plasticity delays
temporal attention
neuromorphic computing
作者简介
张航铭,博士研究生,主要研究领域为脉冲神经网络算法、神经编码,E-mail:zhm0755@tju.edu.cn;白千一,硕士研究生,主要研究领域为脉冲神经元学习算法、人工智能;邓智超,硕士研究生,主要研究领域为类脑计算、神经网络;Alexander Sboev,博士,教授,主要研究领域为神经形态计算、忆阻可塑性;通信作者:于强,博士,教授,主要研究领域为类脑智能计算、神经网络、人工智能、理论神经学,E-mail:yuqiang@tju.edu.cn。
