基于脉冲膨胀可分离注意力与频域学习的双域令牌混合器

打开文本图片集

关键词：脉冲神经网络；脉冲Transformer；膨胀可分离注意力；频域学习；令牌混合器中图分类号：TP183 文献标志码：A 文章编号：1001-3695（2026）04-027-1196-06doi：10.19734/j. issn.1001-3695.2025.08.0303

Spatial-frequency token mixer with spiking dilated separable attention and frequency learning

Tao Huixing，Chen Yunhua，Chen Pinghua，Liao Zhaohui ， Zhou Honghong （School ofComputer Science and Technology，Guangdong University of Technology，Guangzhou51O640，China）

Abstract：Spiking Transformerhasinjectednewvitalityintospiking neuralnetwork（SNN）researchduetotheirpotentialfor energyeficiencyandsuperiorperformance.However，theystillsuerfromquadraticcomputationalcomplexityandunsatisfactory performance.To addresstheseissues，this paper proposedaspatial-frequencytoken mixer （SFTM）that integrated two branches：a spiking dilated separable atention （SDSA）mechanismandaFourier transform-basedfrequencylearner（FTFL）. The SDSA moduleemployed large-kernel depthwiseconvolutions to model long-range dependencies with linear complexity， substantiallyreducingcomputationaloverhead.TheFTF modulerestoredhigh-frequencydetails through spectral transformationandadaptive weighting，efectivelycounteringthelow-assfilteringcharacteristicsofspikingatention mechanismsAdditionall，the paperdesignedaspking depth-wisefeed-forwardnetwork（SDWFFN）tofacilitatecrosschannelinformation exchange.Experimentson CIFAR-1O/10Oandmultipleneuromorphicdatasets demonstrate that the proposed SFformerachieved comparableorsuperioraccuracytostate-of-te-artmodels，whilesignificantlyreducingcomputationalcomplexityandeergyconsumption. Specifically，the model achieved 81.34% accuracy on CIFAR-1OO，outperforming the previous best result by 3. 13% ， with an energy consumption of only 0.27mJ ，validating its significant energy efficiency while maintaining high performance.

Key words：spiking neural network；spiking Transformer；spiking dilated separable attention;frequencylearning；token mixel

0引言

脉冲神经网络（spikingneuralnetwork，SNN）[1～3]是由神经形态计算引入的新型算法范式，凭借其高生物学合理性、事件驱动特性及低功耗优势，SNN被认为是人工神经网络（artificialneuralnetwork，ANN）有竞争力的替代方案。（剩余15903字）