Spiking Neural Networks, the Next Generation of Machine Learning

《Spiking Neural Networks, the Next Generation of Machine Learning》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、ApplausefromLudovicBenistantand56othersDevinSoniFollowcryptomarkets,datascience—100.github.ioJan11·4minreadSpikingNeuralNetworks,theNextGenerationofMachineLearningEveryonewhohasbeenremotelytunedintorecentprogressinmachinelearninghasheardofthecurrent2ndgenerationartic

2、ialneuralnetworksusedformachinelearning.Thesearegenerallyfullyconnected,takeincontinuousvalues,andoutputcontinuousvalues.Althoughtheyhaveallowedustomakebreakthroughprogressinmanyelds,theyarebiologicallyinn-accurateanddonotactuallymimictheactualmechanismsofourbrain’sn

3、eurons.The3rdgenerationofneuralnetworks,spikingneuralnetworks,aimstobridgethegapbetweenneuroscienceandmachinelearning,usingbiologically-realisticmodelsofneuronstocarryoutcomputation.Aspikingneuralnetwork(SNN)isfundamentallydierentfromtheneuralnetworksthatthemachinele

4、arningcommunityknows.SNNsoperateusingspikes,whicharediscreteeventsthattakeplaceatpointsintime,ratherthancontinuousvalues.Theoccurrenceofaspikeisdeterminedbydierentialequationsthatrepresentvariousbiologicalprocesses,themostimportantofwhichisthemembranepotentialofthene

5、uron.Essentially,onceaneuronreachesacertainpotential,itspikes,andthepotentialofthatneuronisreset.ThemostcommonmodelforthisistheLeakyintegrate-and-re(LIF)model.Additionally,SNNsareoftensparselyconnectedandtakeadvantageofspecializednetworktopologies.Dierentialequation

6、formembranepotentialintheLIFmodelMembranepotentialbehaviorduringaspikeSpiketrainsforanetworkof3neuronsAfullspikingneuralnetworkAtrstglance,thismayseemlikeastepbackwards.Wehavemovedfromcontinuousoutputstobinary,andthesespiketrainsarenotveryinterpretable.However,spiket

7、rainsoerusenhancedabilitytoprocessspatio-temporaldata,orinotherwords,real-worldsensorydata.Thespatialaspectreferstothefactthatneuronsareonlyconnectedtoneuronslocaltothem,sotheseinherentlyprocesschunksoftheinputseparately(similartohowaCNNwouldusingalter).Thetemporala

8、spectreferstothefactthatspiketrainsoccurovertime,sowhatweloseinbinaryencoding,wegaininthetemporalinformationofthespikes.This