The Future of Radio Learning Efficient Signal Processing ... · Deep Learning Trends 2 •Large...

Post on 23-Aug-2020

2 views 0 download

Transcript of The Future of Radio Learning Efficient Signal Processing ... · Deep Learning Trends 2 •Large...

hume@vt.eduwww.hume.vt.edu

TheFutureofRadioLearningEfficientSignalProcessingSystems

TimO’Shea,oshea@vt.eduFounder/ChiefScientist,DEEPSIGInc

ResearchFaculty,VirginiaTechhttps://www.deepsig.io/

DeepLearningTrends

2

• LargeNeuralNetworksareDisruptingSignalProcessing

• Biggerchangethanmostpeoplerealize• FeatureLearning• End-to-endlearning• Widelyapplicabletomanydomains

• Featureengineeringisbecomingirrelevant• Experttransformsunnecessarytoachievestateoftheartperformance

• Engineeredfeaturescreatedbarrierstolearninganyway

• Engineeredalgorithmscreatedworkinoptimizingdisparatealgorithms

• Justembraceeverythingasadensemultiplyaccumulatewithsomearbitrarysetofweights

DeepLearningTrends

3

• Thingsthataren’tthatexcitinganymore• ComputerVision(ObjectRecognition)

• Selfdrivingcars(Tesla,Comma.ai,etc)• VoiceRecognition(Siri/GoogleAssistant)

• Itscleardeeplearningcan(has)destroythestateoftheartinthesefields• Customsiliconcandrasticallybringdownpowercostsofthesenetworks• Apple“Bionic”processor,GoogleTPU,etc

• Theseshipsarealreadysailingwholesale

Whatisstillbrewing…

4

• Probablyeverysignalprocessingalgorithmonearthcouldbereconsidered/improved• Especiallylearnedrepresentations…

• Simpleconstructoftheautoencoder• Learnsentirelynewrepresentationsofinformation• Basedonreconstructionlossorotherlossfunctions

Minimizeencoderanddecoderloss

OnlyforrelevantdistributionofX!

What'sinteresting… andcomingnext…

5

• ImageCompressionSchemes• Contextawarecompression:both0.08bits/pixel• NewJPEGStandards

• VideoCompression• Netflix/MITadoptionnow

• Encryption• Simplyamin/maxreconstructionoptimization

BetterEstimatorsunderimpairedchannels

6

• Re-considerestimatorsandrepresentations• Inthecontextofactualdistributioninformation• Betterestimationunderimpairments!• Especiallygoodforshort-timewindows

EfficientApproximateDecoders

7

• Learningdecodersfor‘nearoptimal’errorcorrectioncodes!(Partitioningtoscale)• FECdecoding/detectioncurrentlythe#1powerconsumingoperationinradiobasebanddevices

• WorkfromCammerer,Gruber,Hoydis,tenBrink(UniversityofStuttgart/Nokia-BellLabs)

• Showsnear-optimalpolarcodedecodingperformancewithpartitionedneuralnetworksatlowercomplexitythansuccessivecancellationorBeliefPropagation!

• Potentiallyamajoradvanceinerrorcorrection• Learnapproximatedecodersoncode-wordsets• Lowlatencyone-shotdecodingathigherefficiency

• “ScalingDeepLearning-basedDecodingofPolarCodesviaPartitioning”https://arxiv.org/abs/1702.06901

RethinkApproachtoCommunications

8

“Reproducingatonepointeitherexactlyorapproximatelyamessageselectedatanotherpoint”

• C.E.Shannon,“Amathematicaltheoryofcommunication,”1948

• Allcommunicationsystemsneedtodoisoptimizeforreconstructionloss• Everythingelseisasecondarysub-task• Letsnotgethunguponminutia

• Thisactuallyworksreallywell• Matchescodedmodulationbaselinesimmediately

RadioMethodsforSaliency

9

• Howcanwereducethesearchspace?• Leveragethingsweknowaboutpropagationphysics?

• Introducedomainawareattentionmechanismsintherightway–• Decompositionofreceiver

• Learnedestimationmodules(Attentionmodel)• Experttransformationmodulestomatchphysicalworldpropagationmodels/effects

• Learneddemapping/representationmodules• Jointlearningofencoder/modulator,synchronizer,decoder/demodulator,andovertheairrepresentation• Lowercomplexitylearningproblem

• Convergesfaster,lessoverfitting• Onlyimpartspropagationmediumexpertknowledge

(thingswecan’tchange)• Learneverythingelseend-to-end

Whatifwehavetosharethechannel?

10

• Caneasilyextendthismethodtomulti-accesschannel• Learnabettersolutionthanorthogonality• Samebasicprincipals• Comesupwithimmediatelyinterpretableresults

(a)(1,1)(b)(2,2)(bits,symbols)

SynthesizingComplexMulti-userPHYs

11

• Comparisonwithexistingmethods• Comparethemulti-useraccesschannel• Orthogonal(Time-slicing(TS))vslearnedmethod• LearnsnewneverbeforeseenPHYscheme• Infinitenumberofpossiblewaveforms!• Inthiscasepseudo-orthogonalsuperpositioncode

(a)(1,1)(b)(2,2)(c)(4,4)(d)(4,8)(bits,symbols)

ComplexPHYLearning

SynthesizingComplexMulti-AntennaPHYs

12

• Extendthistechniquetomulti-antenna• Samebasicprincipals• ComplexMIMOchanneleffects

• IncorporateCSIfeedback• EntirelynewMIMOscheme

SynthesizingComplexMulti-AntennaPHYs

13

• Canlearnincrediblycomplexjointsolutions• Softjoint-modulation-codingschemes• Outperformcurrentbaselines(zeroforcingMIMO)

• Enormouspotentialfordistributedwireless• MIMOsystemperformance• Secrecyandprivacy

Non-standardMIMOQAMModesComplexMIMOPHYLearning

Learned2x2Constellations

1. Transmitted2. Diag Rx3. UniformRx

Thanks!Questions?

14

MLDrivenRFSystemsarecomingFASTComeandworkwithus/talktous!

Nextgenerationradiosensingandcommunicationssystems

AppliedResearchMatureCapabilities

UnrestrictedFundamentalEnablingResearch

info@deepsig.io oshea@vt.edu