RR-SC：边缘设备中基于随机计算神经网络的运行时可重配置框架.pdfVIP

计算机研究与发展DOI：10.7544/issn1000-1239.202220738

JournalofComputerResearchandDevelopment61（4）：840−855，2024

RR-SC:边缘设备中基于随机计算神经网络的运行时可重配置框架

宋玉红沙行勉诸葛晴凤许瑞王寒

（华东师范大学计算机科学与技术学院上海200062）

（yhsong@）

RR-SC:Run-TimeReconfigurableFrameworkforStochasticComputing-Based

NeuralNetworksonEdgeDevices

SongYuhong,EdwinHsing-MeanSha,ZhugeQingfeng,XuRui,andWangHan

（SchoolofComputerScienceandTechnology,EastChinaNormalUniversity,Shanghai200062）

AbstractWiththedevelopmentofAIdemocratization,deepneuralnetworks(DNNs)havebeenwidelyappliedto

edgedevices,suchassmartphonesandautomateddriving,etc.Stochasticcomputing(SC)asapromisingtechnique

performsfundamentalmachinelearning(ML)tasksusingsimplelogicgatesinsteadofcomplicatedbinaryarithmetic

circuits.SChasadvantagesoflow-powerandlow-costDNNsexecutiononedgedeviceswithconstrainedresources

(e.g.,energy,computationandmemoryunits,etc.).However,previousSCworkonlydesignsonegroupofsettingfor

fixedhardwareimplementation,ignoringthedynamichardwareresources(e.g.,battery),whichleadstolowhardware

efficiencyandshortbatterylife.Inordertosaveenergyforbattery-powerededgedevices,dynamicvoltageand

frequencyscaling(DVFS)techniqueiswidelyusedforhardwarereconfigurationtoprolongbatterylife.Inthispaper,

wecreativelyproposearun-timereconfigurableframework,namelyRR-SC,forSC-basedDNNsandfirstattemptto

combinehardwareandsoftwarereconfigurationstosatisfythetimeconstraintofinferenceandmaximallysave

energy.RR-SCusingreinforcementlearning(RL)cangeneratemultiplegr