programmable nanophotonics for quantum information … · 2020-03-23 · programmable nanophotonics...
TRANSCRIPT
ProgrammableNanophotonicsforQuantumInformationProcessing and
Artificial Intelligence
DariusBunandar,NicholasHarris, DirkEnglundQuantumPhotonics Lab
CadencePhotonicsSummitSept.6th 2017
Acknowledgments
Funding
Prof.DirkEnglund (MIT)
Dr.NickHarris (MIT)
Prof.SethLloyd (MIT) Prof.MarinSoljačić (MIT)
Dr.TomBaehr-Jones (Elenion) Dr.MichaelHochberg (Elenion)
Outline
1. Programmablenanophotonic processor2. Photonicquantuminformation processing
3. Opticalneural network
Programmablenanophotonic processor
660µm
88MZIs,26inputmodes,26outputmodes,176phase shifters
ImagecourtesyofAFRL Rome
Thearrayin action
Stronglaserinput
Unitcell performance
>70dB visibility
Quantumsimulator architecture
Single-photonsources
Linearoptics unitary Single-photondetection
MichaelReck,AntonZeilinger,HerbertJ.Bernstein,andPhilipBertaniPhys. Rev. Lett. 73,58(1994)WRClements,PC Humphreys,B J Metcalf,WSKolthammer,IAWalmsleyOptica 3(12),1460-1465 (2016)
Unitcellperformanceasa qubit
Dualrail encoding
Randomized benchmarking
Photosynthesisandquantum transport
Environment-assistedquantumtransport (ENAQT)
P.Rebentrost,M.Mohseni,I.Kassal,S.Lloyd,A.Aspuru-Guzik,New Journal of Physics 11033003 (2009)
Noisyscattering simulations
Fully-integratedphotonicquantum computer
Heraldedsingle-photon sources
400µm
NicholasC.Harris,DavideGrassani,AngelicaSimbula,MihirPant,MatteoGalli,TomBaehr-Jones,MichaelHochberg,DirkEnglund,DanieleBajoni,ChristopheGallandPhys. Rev. X 4041047 (2014)
On-chipsingle-photon detectors
FarazNajafi*,JacobMower*,NicholasC Harris*,FrancescoBellei,AndrewDane,CatherineLee,XiaolongHu,PrashantaKharel,FrancescoMarsili,SolomonAssefa,KarlKBerggren,DirkEnglundNature Comm. 65873 (2015)
Incollaboration withProf.KarlBerggrenat MIT
Deeplearning
Artificialneural network
Canitbedonewith light?
Singularvalue decomposition
Opticalneural network
Vowelrecognition task
Powerinlog-spacedfreq. bands
90people speak4 vowels
Fouriertransform
360samples:180training+180 test
Thecaseforopticalneural network● Fast,low-energymatrixcomputation● Lowthermalnoise:goodforanalog encoding● NN’smoreresilienttoerrorsthangeneral-purpose computer
Equivalentcomputeperformanceof ONN:
R = m ×N2 ×BW FLOPS;m=layers,N xN matrixmultiplication,BW=bandwidth(>10 GHz)
Energy/FLOP Error propagation? Classification error
DigitalElectronic (GPU) ~100pJ/FLOP*, includingmemory retrieval
zero Low
Optical NN ~10/NfJ(signal re-gen)~10/(m×N)fJ (all-optical)
Betterthan10-bitprecisionforN=4096 with16-bitphase settings
Low(atleastforlowN<4096)
*M.Horowitz,Solid-StateCircuitsConferenceDigestofTechnicalPapers(ISSCC),2014IEEEInternational,10–14. IEEE.
Summary
1. Quantum simulation
2. Opticalneuralnetwork:potentialfornearlyenergy-freematrix multiplication
Outlook
● CMOSandphotonic integration
● Novelquantumphotonic devices○ Single-photon sources○ Single-photondetectors:Ge APDs○ MEMS integration High-Q
Output
Store-and-releasesinglephoton source
MikkelHeuck,MihirPant,DirkEnglund arXiv:1708.08875TaeJoonSeok,NielsQuack,SangyoonHan,RichardMuller,MingWuOptica 3(1)64-70 (2016)