11.Make Your Own Neural Network.zip

Contents





Prologue

TheSearchforIntelligentMachines

ANatureInspiredNewGoldenAge





Introduction

Part1HowTheyWork

EasyforMe,HardforYou

ASimplePredictingMachine

ClassifyingisNotVeryDifferentfromPredicting

TrainingASimpleClassifier

SometimesOneClassifierIsNotEnough

Neurons,Nature’sComputingMachines

FollowingSignalsThroughANeuralNetwork

MatrixMultiplicationisUseful..Honest!

AThreeLayerExamplewithMatrixMultiplication

PreparingData





Part2DIYwithPython

Python

InteractivePython=IPython

AVeryGentleStartwithPython

NeuralNetworkwithPython

TheMNISTDatasetofHandwrittenNumbers



CreatingNewTrainingData:Rotations





Epilogue





AppendixA:

AGentleIntroductiontoCalculus

AFlatLine

ASlopedStraightLine

YoucandoCalculus!





AppendixB:

DoItwithaRaspberryPi

InstallingIPython

MakingSureThingsWork

TrainingAndTestingANeuralNetwork

RaspberryPiSuccess!



Prologue

We’venotbeensatisfiedbymechanicalorelectronicmachineshelpinguswithsimpletasks

flintsparkingfires,pulleysliftingheavyrocks,andcalculatorsdoingarithmetic.



Instead,wewanttoautomatemorechallengingandcomplextaskslikegroupingsimilarphotos,

recognisingdiseasedcellsfromhealthyones,andevenputtingupadecentgameofchess.

Thesetasksseemtorequirehumanintelligence,oratleastamoremysteriousdeepercapability

ofthehumanmindnotfoundinsimplemachineslikecalculators.



Machineswiththishumanlikeintelligenceissuchaseductiveandpowerfulideathatourculture

isfulloffantasies,andfears,aboutittheimmenselycapablebutultimatelymenacingHAL



Sostrongisourdesireforintelligentmachinesthatsomehavefallenforthetemptationtocheat.

TheinfamousmechanicalTurkchessmachinewasmerelyahiddenpersoninsideacabinet!







ANatureInspiredNewGoldenAge

inthe1950s.Initialsuccessessawcomputersplayingsimplegamesandprovingtheorems.

Somewereconvincedmachineswithhumanlevelintelligencewouldappearwithinadecadeor

so.



Butartificialintelligenceprovedhard,andprogressstalled.The1970ssawadevastating

academicchallengetotheambitionsforartificialintelligence,followedbyfundingcutsanda

machineintelligenceoutofitsrut.Whynottrytobuildartificialbrainsbycopyinghowreal

biologicalbrainsworked?Realbrainswithneuronsinsteadoflogicgates,softermoreorganic

reasoninginsteadofthecoldhard,blackandwhite,absolutisttraditionalalgorithms.



Scientistwereinspiredbytheapparentsimplicityofabeeorpigeon'sbraincomparedtothe

complextaskstheycoulddo.Brainsafractionofagramseemedabletodothingslikesteer

flightandadapttowind,identifyfoodandpredators,andquicklydecidewhethertofightor



Butwithtraditionalapproachestosolvingproblemsthesecomputerswithmassivestorageand

superfastprocessorscouldn’tachievewhattherelativelyminisculebrainsinbirdsandbees

coulddo.



Neuralnetworksemergedfromthisdriveforbiologicallyinspiredintelligentcomputingand

wentontobecomeoneofthemostpowerfulandusefulmethodsinthefieldofartificial

intelligence.Today,Google’sDeepmind,whichachievesfantasticthingslikelearningtoplay

videogamesbyitself,andforthefirsttimebeatingaworldmasterattheincrediblyrichgameof

Go,haveneuralnetworksattheirfoundation.Neuralnetworksarealreadyattheheartof





 