Convolutional Networks for Fast, Energy-Efficient

《Convolutional Networks for Fast, Energy-Efficient》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、ConvolutionalNetworksforFast,Energy-EcientNeuromorphicComputingStevenK.Esser,PaulA.Merolla,JohnV.Arthur,AndrewS.Cassidy,RathinakumarAppuswamy,AlexanderAndreopoulos,DavidJ.Berg,JereyL.McKinstry,TimothyMelano,DavisR.Barch,CarmelodiNolfo,Pallab

2、Datta,ArnonAmir,BrianTaba,MyronD.Flickner,andDharmendraS.ModhaIBMResearch{AlmadenAbstractmentalchangestonetworkstate,bothformallyrequiredforDeepnetworksarenowabletoachievehuman-levelperformanceonbackpropagation-basedgradientlearning.Incomparison,neu-ab

3、roadspectrumofrecognitiontasks.Independently,neuromorphicromorphicdesignsuse1-bitspikestoprovideevent-basedcom-computinghasnowdemonstratedunprecedentedenergy-eciencyputationandcommunication(consumingenergyonlywhenthroughanewchiparchitecturebasedonspikingn

4、eurons,lowpre-necessary)anduselow-precisionsynapsestoco-locatememorycisionsynapses,andascalablecommunicationnetwork.Here,wewithcomputation(keepingdatamovementlocalandavoidingdemonstratethatneuromorphiccomputing,despiteitsnovelarchi-o-chipmemorybottlenecks

5、).Here,wedemonstratethatbytecturalprimitives,canimplementdeepconvolutionnetworksthatintroducingtwoconstraintsintothelearningrule{binary-i)approachstate-of-the-artclassicationaccuracyacross8stan-darddatasets,encompassingvisionandspeech,ii)performinferencev

6、aluedneuronswithapproximatederivativesandtrinary-whilepreservingthehardware'sunderlyingenergy-eciencyandhighvalued(f1;0;1g)synapses{itispossibletoadaptbackprop-throughput,runningontheaforementioneddatasetsatbetween1200agationtocreatenetworksdirectlyimple

7、mentableusingen-and2600framespersecondandusingbetween25and275mW(ef-ergyecientneuromorphicdynamics.Thisdrawsinspirationfectively>6000frames/sec/W)andiii)canbespeciedandfromthespikingneuronsandlow-precisionsynapsesofthetrainedusingbackpropagationwiththesam

8、eease-of-useascontem-brain[10],andbuildsuponworkshowingthatdeeplearningporarydeeplearning.Forthersttime,thealgorithmicpowerofcancreatenetworkswithconstrainedconnectivity[11],low-deeplearningcanbemergedwithth