PhysicalNeuralNetworks JonathanLamont November 16,2015 Contents • • • • ArtificialNeuralNetwork HistoryofDevelopment AHaH Computation FutureGoalsofAHaH System NeuralComputation • Understandingbiologicalneuralsystemsfalls inaspectrumbetweenmimicryand algorithmicsolutions • Tomimicthebrain,shiftquestionfrom“how dobrainscompute?”to“howdobrainsbuild andrepairthemselvesasdissipativeattractorbasedstructures?” WhatisanArtificialNeuralNetwork? • ANNsaregenerallydenselyconnectedgroups ofindividualcells,calledneurons,that producedesiredoutputsfortrainedinputs (trainedratherthanprogrammed) • Inspiredbybiologicalneuralnetworks • Usedinmachinelearningandcognitive sciencetoestimatefunctions NeuronDiagram ANNDiagram InspirationforPhysical Implementation • IBM’sSimulationwith1BillionNeuronsand10 TrillionSynapsesrequired147,456CPUs,144TBof memoryrunning,andrequires2.9MWsofpowerat 1/83real-time. – Wouldrequire7GWofpowertosimulatehumancortexbasedon theseestimates, usingvonNeumannArchitecture – Averagehumanneocortex contains150,000billionconnections • CoreAdaptivePowerProblem:Energywastedduring memory-processorcommunication. AdaptivePowerProblem • Constantdissipationoffreeenergyallowsliving systemstoadaptatallscales • Eachadaptationmustreducetomemory-processor communicationasstatevariablesaremodified – Energyconsumedinmovingthisinformationgrows linearlywithnumberofstatevariablesthatmustbe continuouslymodified HistoryofDevelopment • AlanTuring’s‘IntelligentMachinery’(1948) – Describes machineconsistingofartificialneuronsconnectedinany patternwithmodifierdevices – Modifierdevices couldpassordestroyasignal • Hebb proposedsynapticplasticity(1949) – Describestheabilityofsynapsestostrengthorweaken over timeinresponsetoincreasesordecreasesinactivity • ADALINEor“AdaptiveLinearNeuron”(1960) – Physicaldeviceusingelectrochemicalplatingofcarbonrodsto emulatesynapticelementscalledmemistors. – Memistor – Resistorwithmemory abletoperformlogic operationsandstoreinformation. HistoryofDevelopment(Continued) • Memristor Theory(1971)– regulatesflowof electricalcurrentincircuitandrememberscharge thatpreviouslyflowed;retainmemorywithout power(non-volatile) • Very-large-scaleintegration(VLSI)(1980)–integrated circuitcreatedbycombiningthousandsoftransistors intosinglechip – HelpedNeuralNetworks(Hopfield),Neuromorphic Engineering(Mead),andPhysicsofComputation (Feynman) AHaH Computation • Exploitcontroversial“FourthLawof Thermodynamics”toretrainweightsfrom Swenson(1989): – “Asystemwillselectthepathorassemblyofpathsout ofavailablepathsthatminimizesthepotentialor maximizestheentropyatthefastestrategiventhe constraints.” • Thisthermodynamicprocesscanbeunderstood: 1)Particlesspreadoutalongavailablepathwaysanderode differentials thatfavoronepathortheother 2)Pathwaysthatleadtodissipationarestabilized Figure1.AHaHprocess. NugentMA,MolterTW(2014)AHaHComputing–FromMetastableSwitchestoAttractorstoMachineLearning.PLoSONE9(2):e85175. doi:10.1371/journal.pone.0085175 http://journals.plos.org/plosone/article?id=info:doi/10.1371/journal.pone.0085175 AHaH Plasticity • Definesynapticweight: w=Ga – Gb,whereGa andGb areconductancevariables • Anti-Hebbian (erasepath) – Modificationtosynapticweightthatreducesprobability thatthesynapticstatewillremainthesameupon subsequent measurement • Hebbian (selectpath) – Modificationtosynapticweightthatincreasesprobability thatsynapticstatewillremainthesameuponsubsequent measurement Figure4.Adifferentialpairofmemristorsformsasynapse. NugentMA,MolterTW(2014)AHaHComputing–FromMetastableSwitchestoAttractorstoMachineLearning.PLoSONE9(2):e85175. doi:10.1371/journal.pone.0085175 http://journals.plos.org/plosone/article?id=info:doi/10.1371/journal.pone.0085175 LinearNeuronModel • y=b+ΣNi=0xiwi – yisthe output ofthe selected neuron,bisthe bias to the function,xi isthe output ofneuron i,and wi isthe weight between the selected neuron and neuron xi • Weights and biases change based onthe AHaH rule – αand βare constants that control the relative contribution ofAnti-Hebbian and Hebbian plasticity AHaH forMachineLearning • AHaH node:usesAHaH ruletobisectinput spaceusingahyperplane tomakeabinary decision • SincetherearemanywaystheAHaH nodecan bisectinput,thehyperplane chosentobisect theinputspaceis“theonethatmaximizesthe separationofthetwoclasses”ofinputs. Figure2.Attractorstatesofatwo-inputAHaHnode. NugentMA,MolterTW(2014)AHaHComputing–FromMetastableSwitchestoAttractorstoMachineLearning.PLoSONE9(2):e85175. doi:10.1371/journal.pone.0085175 http://journals.plos.org/plosone/article?id=info:doi/10.1371/journal.pone.0085175 AHaH PowerConsumption • MajormotivationistoeliminatevonNeumann bottleneckregardingmemory-processor communication • Energyoutputisdefined: • Inpractice,theenergyconsumptionpersynapse perupdateisaround12pJ,althoughscaling trendsprojectenergyconsumptiontobelowered to2pJ inthefuture. FutureGoals • Provideaphysicaladaptivelearninghardware resource(AHaH circuit)thatprovidesmemory similartohowRAMprovidesmemoryto currentcomputingsystems. • UseAHaH nodeasabuildingblockforhigherorderadaptivealgorithmsnotyetconceived. Conclusion • ArtificialNeuralNetworksaregreatfunction approximations,butthereisamajorbottleneck frommemory-processorcommunicationusing vonNeumannarchitecture • Reachingbacktothelate1940s,thistypeof computationhasbeentheorizedandstudied • AHaH computationexploitsphysical phenomenontoimplementanartificialneural network,lesseningthememory-processor communicationbottleneckforpowerwhile allowingmassiveparallelism. Sources • AHaH Computing-FromMetastableSwitchestoAttractorsto MachineLearning – http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0085175 • “NeuralNetworkLearning” – http://web.eecs.utk.edu/~mclennan/Classes/420-527-S14/handouts/Part-2A1pp.pdf • “Neuralnetworkmodelsforpatternrecognitionand associativememory” – http://www.sciencedirect.com/science/article/pii/089360808990035X • “Perceptrons,Adalines,andBackpropagation” – http://www-isl.stanford.edu/~widrow/papers/bc1995perceptronsadalines.pdf