brain inspired compung - kirchhoff institut für physikmeierk/myfiles/downloads/meierlyonict… ·...

34
Brain Inspired Compu0ng Karlheinz Meier Ruprecht‐Karls‐Universität Heidelberg FACETS Project Coordinator ICT Conference, Lyon (France), 2008

Upload: others

Post on 08-Jun-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

BrainInspiredCompu0ngKarlheinzMeierRuprecht‐Karls‐UniversitätHeidelberg

FACETSProjectCoordinator

ICTConference,Lyon(France),2008

Page 2: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

FutureEmergentTechnologies(FET)inIST

„FETwillexploreradicalinterdisciplinaryavenues,deliveringproofs‐of‐conceptfornewop;onsanddemonstra;ngnewpossibili;es.ItwillstrengthenEurope'sscienceandtechnologybaseinnewandemergingareas,refinenewvisionstothepointwheretheyaCractindustrialinvestment,andestablishnewinterdisciplinaryresearchcommuni;eswithinEuropeanscienceandindustry“

Especially:

“RecentadvancesinICTandneuroscienceenableasignificantpartofthehumanbraintobestudiedandmodelledin‐silico.Thisobjec;veseekstoexploitsuchadvancesinordertobeCerunderstandhowthebrainprocessesinforma;onand/orhowitcommunicateswiththeperipheralnervoussystem(PNS),andtoexplorepoten;alapplica;onsofthis”

Page 3: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

„Complicated“and......„Complex“

Top‐DownDesign BoVom‐upEvolu0on

Page 4: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

1  WhatistheUniversemadeof?

2  WhatistheBiologicalBasisofConsciousness?

3  WhydoHumansHaveSoFewGenes?

4  ToWhatExtentAreGene0cVaria0onandPersonalHealthLinked?

5  CantheLawsofPhysicsbeUnified?

.....120morefollowing....

125Scien0ficKeyQues0onsiden0fiedbythe

„AmericanAssocia;onfortheAdvancementofScience“

WHATDON‘TWEKNOW...

Page 5: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

ContemporaryITsystems Processor‐memorybasedarchitectureswithserialcommandexecu0on(Turing)

 Predeterminedalgorithmsdefinecapabiliesandperformance(sobware)

 Basedonwelldefinedreproduciblestatesandwelldefinedreversible0meevolu0on

 Electronicsimplementa0onofBooleanoperators,highpowerconsump0on

 Extremelyhighyieldrequirements,liVlefaulttolerance

 Limitedbyatomicdistancescaleincomponets(nm):componentlimited

WELLUNDERSTOOD

NeuralcomputaCon Maximallyparallel,non‐linearcomp0ngelementswithlargediversity

 Timecorrela0onsdrivethedynamics

 Learningbyinternalself‐organisa0onandstronginterac0onwithenvironment

 Lowpowerconsump0onandhighfaulttolerance

 Limitedbydegreeofcomplexity:architecturelimited

NOTUNDERSTOOD(listedasamajorchallengefor21.Centuryscience)

Page 6: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von
Page 7: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von
Page 8: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

RealBiology:Neurons‐Synapses‐Dendrites‐Spikes

Ac0onPoten0al=Spike

Page 9: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

I(t) = u(t)/R + C⋅du/dt u(t) : membrane potential I(t) : input current

Using RC time-constant :

τm ⋅ du/dt = - u(t) + R⋅ I(t)

In addition :

„Spike-Generation“ and „Reset“ of u if u = ϑ from:W.Gerstner,SpikingNeurons

Example for a Neural Circuit - here : „Integrate and Fire“!A simple approch to neuromorphic electronics!

Page 10: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

Why neuromorphic electronics ?!TWO Answers : !

I.  Research tool for neuroscience : Study neural architectures in „ANY“ detail at „ANY“ speed !

II.  New type of information processing : Use massively parallel (10LARGE), potentially low yield (solve fabrication problem), low power (solve leakage current problem). Make use of self-organisation (solve software problem).!

Rela0veweigh0ng?DoesII.dependonI.?Interdisciplinary!

Page 11: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

StatusandFutureofNeuromorphicCompu0ng

• Humanlevelperformance• Dawnofanewage

vonNeumannmachines

NeuromorphicMachines

MachineComplexitye.g.Gates;Memory;Neurons;SynapsesPower;

Size

Dawnofanewparadigm

“simple” “complex”

EnvironmentalComplexitye.g.InputCombinatorics

[log]

ProgramObjec0ve

Atradebetweenuniversalityandefficiency

TakenfromtheDARPASynapseProjectCall

Weareheretoday:Machinecomplexityofneuromorphicdevicess0llexceedstheoneofvonNeumannmachines.Currentproblemstackledwithneuromorphicdevicesares0llsimple.Newcomputa0onalparadigmsareexpectedpayoffwithwithmorecomplexinputdataandlargernetworks.

Page 12: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

ConceptualapproachestoNeuromorphicCompu0ng

 “BoVom‐up”:Basedon(simplified)reconstructedmodelsofbiologicalmorphologyandfunc0onality:cells,synapses,connec0vity,plas0citymechanisms(“atomsandtheirinterac0ons”)(e.g.FACETSproject)

 “Top‐down”:Basedonlargescalefunc0onalblocksandtheirconnec0vity(e.g.JeffHawkins)

 “Firstprinciples”:Designofcomputa0onalparadigms,notnecessarilybiologicalplausible(e.g.FACETS,liquidcompu0ng)

 “Evolu0onary”designofcomputa0onalstructures

Page 13: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

Speedw.r.t.biologicalreal‐0me10‐3

101

10‐2 10‐1 105

mm

3 ofb

rain0ssue

sim

ulated

orem

ulated

1mm

3 =105neu

ronsand

109synapses

100

103

102

101

FACETSVLSIApproach(upto108neurons)

12.000nodeCluster

40nodePIV

16nodePIIIBiologicalreal‐0m

e

104

Neocortex(1012neurons)

V1(109neurons)

FACETS:ComplementaritySupercomputersandVLSI‐Complexityvs.Speed

10%V1(108neurons)

Page 14: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

TheMeritsoffast(105)neuralVLSI

Acess<10OrdersofMagnitudeinTimeinanar0ficialSystemwithaspa0alcomplexityof>105?

Page 15: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

The FACETS

Con

sorti

um Fast Analog Computing with

Emergent Transient States

U Bordeaux, CNRS (Gif-sur-Yvette and Marseille), U Debrecen, TU Dresden, U Freiburg, TU Graz, U Heidelberg *, EPFL

Lausanne, Funetics S.a.r.l. Lausanne, U London, U Plymouth, INRIA Sophia-Antipolis, KTH Stockholm

*Coordinator

An Integrated Project in the 6th Framework Programme

Information Society Technology - Future Emergent Technologies

FP6-2004-IST-FETPI Project Reference 15879

Page 16: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

FACETS:BasicIdea,methodologicalapproachandgoalsNeurobiology:StructuralandFunc0onalInves0ga0onofthe

Neocor0calMicrocircuitandtheCircuitElementsin‐vivoandin‐vitro

Modelling:VirtualMicrocircuitsonState‐of‐the‐ArtComputers

Hardware:Emula0oninanalogandmixed‐signalVLSIsystems

CommonGoal:Studynon‐classicaluniversalcompu0ngsolu0ons

Benchmarking(biologyvs.Modellingvs.HardwarewithvisualtasksinVI)

Methodology:ToolDevelopment(Compu0ng,VLSI)ReducConofBiologicalDetail/Complexity

Page 17: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

synapses:

pk,l(t) exponen0alonsetanddecay(spikeshape)gk,l 0togmaxwith4bitresolu0on

effec0vemembrane0me‐constantcm /gtotalis0me‐dependent

high‐conductancestatepossible

current source, no voltage dependence

A FACETS Approach : Conductance-based Network Model

sumoverexcitatorysynapsecurrentsk

sumoverinhibitorysynapsecurrentsl

Voltage dependent part, changes membrane conductance

membranecurrent

leakagecurrent

( ) ( ) ( )∑∑ −+−+−=l llk kk EVgpEVgpEVg

dtdVc ixlleakm

Page 18: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

FACETSmixed‐signalVLSISystemStage1(Chipbased)

100.000dynamicSynapses(PlasCcity)

384Neurons

DigitalControl

Inherentlyfast:approx.1050mesbiology

Page 19: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

Upto16neuralnetworkboards(“crate”)canbecombinedtobuildalargernetwork

Page 20: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

FACETSHardware

NESTSoftware

100ms

2,5µs

Equivalence between Hardware and Software!Subthreshold Membrane Potential and Spiking

Page 21: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von
Page 22: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

WaferScaleIntegra0on:Spa0allySeparatedFunc0ons:NeuralVLSI‐ContemporaryIT

Wafer

label0mestamplabel

0mestamplabel

0mestamp

packettransportFPGA

PCB

Layer2

Layer1

HICANN

HierarchicalSetup

•  Layer1:Massivelyparallelneuralcircuit,usecompletewafertoexploitinherentfaulttolerance,asynchronouspulsetransmission

•  Layer2:Packetbaseddigitalcommunica0onprotocolforlong‐rangeSpiketransmission,networkset‐up,monitoringandcontrol.

Page 23: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

J.Fieresetal.,IJCNN2008

NewTechnologyexploringtheinherentfaulttolerancsofneuralcurcuits:WaferScaleIntegra0on

Page 24: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

J.Fieresetal.,IJCNN2008

FACETS:Wafer‐ScaleIntegra0on

Page 25: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

J.Fieresetal.,IJCNN2008

Page 26: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

J.Fieresetal.,IJCNN2008

Page 27: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

SystemunderConstruc0oninHeidelberg

Page 28: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

TechnologicalChallengesandConstraintsofCMOSNC

  Automated design and verifica0on technologies for verylarge scale (50M synapses per wafer) massively parallel VLSIstructures

  Distributed compact on‐chip / on‐wafer memorytechnologies for parameter storage andplas0city / learning /adapta0on mechanisms (SRAM, current memories, floa0nggates),NEW:non‐CMOSpostprocessing:magne0cstructures

  On‐wafer / inter‐die horizontal high density connec0ontechnologies

 Off‐waferver0calhighdensityconnec0ontechnologies

 NEW : Inter‐wafer 3‐dimensional connec0on technologies(waferstacking)withlowpoweranalogdesign

 NEW : Access to deep sub‐micron (< 100 nm) full waferproduc0onforhigherintegra0ondensi0esofneuralcells

 Exploita0onoftheintrinsicfault‐andmismatch‐toleranceofneuralcircuits

Page 29: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

Interna0onalTechnologyRoadmapforSemiconductors(ITRS)

Page 30: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

Faults??

Page 31: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

CMOL(CMOSMolecularHybrid)(Hammerström,Likharev)

 UseanACTIVECMOSsubstructure(asadvancedaspossible/economical)

 UsePASSIVEnanoscaleconnec0ngelements(0ps+wires)

 UseACTIVEtwo‐terminalnanodevices(e.g.molecules)linkingtheconnec0ngelements

Page 32: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

FromK.LikharevStonyBrook

Page 33: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

LogNneuron

3

4

5

6

7

8

9

10

2008 2009 20122010 20132011

Singlechip189nm

Crate180nm

Singlewafer180nm

Mul0wafer180nm

Mul0wafer180nm

Singlewafer<180nm

WaferStack<180nm

Mul0wafer<180nm

Mul0WaferStack<180nm

DARPA„CatChallenge“

DARPA„MouseChallenge“

FACETS‐1 Possiblefutureroadmap

Page 34: Brain Inspired Compung - Kirchhoff Institut Für Physikmeierk/myfiles/downloads/MeierLyonICT… · here today : Machine complexity of neuromorphic devices sll exceeds the one of von

Bio‐ICT:Excellentopportunitytoini0atethenextHARDWARErevolu0onininforma0ontechnology

Requires:

 Massivelyinterdisciplinaryapproach(includinggraduatestudenttraining) Systema0cefforttowardsbrainmapping(morphologyandfunc0on) Theoryefforttowardscomputa0onalprinciplesandphysicsofcomplexsystems Infrastructureandcapabilityforlargescalehardwaresystemdevelopment Accesstodeep‐submicrontechnologiesandcuzng‐edgeconnec0ontechnologies

Mul0‐ScaleFundingapproach:

Concepts: AdressedinIntegratedProjects(WEAREHERE) Technologies: AdressedinDesignStudies Prototypes: AdressedinPreparatoryPhases Systems: Adressedinjointprojectswithindustry

Europeisinanexcellentsitua0on,leadingthefield!