Transcript
Page 1: Applying Techniques in Supervised Deep Learning to ...cs229.stanford.edu/proj2016/poster/SriramPenkovYe...benefits (between 0.001 and 0.007 RMSE on test sets). • RMSE of the above

Printing:Thisposteris48”wideby36”high.It’sdesignedtobeprintedonalargeformatprinter.

CustomizingtheContent:Theplaceholdersinthisformattedforyou.placeholderstoaddtext,orclickanicontoaddatable,chart,SmartArtgraphic,pictureormultimediafile.

Ttext,justclicktheBulletsbuttonontheHometab.

Ifyouneedmoreplaceholdersfortitles,contentofwhatyouneedanddragitintoplace.PowerPoint’sSmartGuideswillhelpyoualignitwitheverythingelse.

Wanttouseyourownpicturesinsteadofours?Noproblem!JustpictureMaintaintheproportionofpicturesasyouresizebydraggingacorner.

Applying Techniques in Supervised Deep Learning to Steering Angle Prediction in Autonomous Vehicles

Vinay Sriram, Petar Penkov, James Ye

PROBLEM DEFINITION

IMAGE PRE-PROCESSING

CONVOLUTIONAL NEURAL NETWORK MODEL

RESULTS FUTURE WORK

• [1]Eliminatetop-halfoftheimages,anddownsamplebyafactorof100.Thissizecapturesnecessaryinformationforpredictionandisoflowenoughdimensionalitytoallowtraininganetworktobecomputationallyfeasible.

• [2]Applyedge-detection:Constructanimage𝑋 basedonthresholdingandtheSobelkerneloperatoroninput𝑋’:

Define 𝑆), 𝑆+ =−1 0 1−2 0 2−1 0 1

∗ 𝑋′,1 2 10 0 0−1 −2 −1

∗ 𝑋′

• Thedevelopmentofeffectiveautonomousvehiclesisapopularapplicationofmachinelearningandcontrol.

• Wewishtosolvealearningproblempredictingsteeringangles 𝛼4…𝛼6 7 overthecourseofaroadsegmentfromUdacity’s low-resolutionimages 𝑋4 …𝑋6 7 whereeach𝑋8 isdefinedbya640×480×3 RGBtensor.

• Wepartitionthedrivingdatawitha70:30train-to-testratiosuchthatthetest-setcontainssignificantturning.

• Introducemoreclasslabelstobetterapproximateanglecontinuity(forthis,moredatamustbecollectedtoprovideasufficienttrainingvolumeforhigher-magnitudelabels).

• Introduceadditionalconv.layerstoreducethebiasofthemodel,andrunstochasticgradientdescent(onresourceswithmorecomputationalpower)foragreaternumberofepochsforbetterconvergenceatlocaloptima.

• Computecross-imagegradientsthatcandeterminedirectionsoffeaturechangetobetterpredictturnangles.Foreach(𝑖, 𝑗) intheoutputimage,set𝑋8B = 255 ifeither

ofthefollowingaretrue,andset𝑋8B = 0 otherwise.• Sobelgradientmagnitude 𝑆)(𝑖, 𝑗)D + 𝑆+(𝑖, 𝑗)D isabove

somecutoffthreshold(preservingonlytheedges).• Grayscalevalueofthepixel(𝑖, 𝑗) isabovesomecutoff

threshold(preservingonlywhiteornear-whitesectionsoftheimage).Together,lanecaptureisreasonable.

𝑋′ 𝑋

Wediscretizetheproblem:eachsteeringanglegetsaclasslabel𝑥 ∈ 1,101 ,whereeachlabelrepresentsarangeof~0.01rad.ForwardPropagation:𝑧 IJ4 = 𝑊(I)𝑎(I) and𝑎 IJ4 = 𝑓I(𝑧 I )with𝑎8

(4) = 𝑥8 andℎO,P = 𝑎(QJ4).Layer𝑙 hasactivationfunc.𝑓(I) .

MaxPooling

[ClassProbabilities]

[InputImage]

[SoftmaxLayer] [Fully-ConnectedLayer] [Conv.Layer- ReLuActivation]

REFERENCESCameron,Oliver.OpenSourcing223GBofDrivingData.Medium,5Oct.‘16.

Pomerleau,Dean.NeuralNetPerceptionforMobileRobotGuidance.(1993)

UnsupervisedFeatureLearningandDeepLearningTutorial.Stanford.Web.http://ufldl.stanford.edu/tutorial/supervised/MultiLayerNeuralNetworks

Cost[𝑚 examples]:𝐽 𝑊, 𝑏 = 46∑ 4

DℎO,P 𝑥8 − 𝑦8

D68X4

TrainingusestheBack-Propagation algorithmtooptimizethecostJ.Theupdateruleis:∆𝑊 I ∶= ∆𝑊 I + 𝛻O \ 𝐽

with𝑊 I = 𝑊 I − 𝛼 46∆𝑊 I .

𝛿 I = 𝑊 I 7𝛿 IJ4 𝑓′(𝑧 I )

𝛻O \ 𝐽 = 𝛿(IJ4) 𝑎 I 7

𝛿^I = 𝑢𝑝𝑠𝑎𝑚𝑝𝑙𝑒(𝑊 I 7𝛿^

IJ4 )𝑓′(𝑧^I )

𝛻Oc\ 𝐽 = ∑ 𝑎 I ∗ 𝑟𝑜𝑡(𝛿^

IJ4 )68X4

ForTypicalLayers:

ForConvolutionalLayers

• Fortrainingandtesting,weconsideredinparticularhighwaydriving.Theconv.neuralnetworkwaseffectiveatpredictingsteeringanglewithinreasonableerror.

• Sobelpre-filteringofimagesprovidedverymarginalbenefits(between0.001and0.007RMSEontestsets).

• RMSEoftheabovetestsegmentwas0.031(withoutSobelpreprocessing)and0.034(withpreprocessing).

Top Related