Ciencias de Datos y Deep Learning: Neuronas artificiales ... · Ciencia de Datos y Deep Learning:...

Post on 21-Jul-2020

5 views 0 download

Transcript of Ciencias de Datos y Deep Learning: Neuronas artificiales ... · Ciencia de Datos y Deep Learning:...

CienciasdeDatosyDeepLearning:Neuronasartificialesparaaprender

Francisco Herrera

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

2

Historyof

DataScience

BigDataDeepLearning

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

3

En1943WarrenMcCullochyWalterPitts presentaronsumodelodeneuronaartificial,ydescribieronlosprimerosfundamentosdeloquesellamaríaposteriormenteredesneuronales.

W.McCullochandW.Pitts(1943).ALogicalCalculusofideasImmanentinNervousActivity.BulletinofMathematicalBiophysics5:115-133.

RedesNeuronales

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

4

RedesNeuronales

Unaredneuronalseproponecomounsistemainteligentequeimitaalsistemanerviosoyalaestructuradelcerebro,peromuydiferenteentérminosdesuestructurayescala.Aligualquelasneuronasbiológicas,lasneuronasartificialesseinterconectanparaformarredesdeneuronasartificiales.Cadaneuronaartificialutilizaunafunciónprocesamiento queagregalainformacióndeconexionesdeentradaconotrasneuronalesartificiales,unafuncióndeactivaciónyunafuncióndetransferenciaparadarunasalidadelaneuronaensusconexionesdesalida.

Netj

x1

x2

xn

wj1

wj2

wjn

yj å=

×=×++×+×=n

iijinjnjjj xwxwxwxwNet

12211 !

÷÷ø

öççè

æ-×=-= å qq ijijjj xwfNetfy )(

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

5

RedNeuronalClásica:Backpropagation

Credits:TheEvolutionofNeuralLearningSystems:ANovelArchitectureCombiningtheStrengthsofNTs,CNNs,andELMs.NMartinel, CMicheloni…- IEEESMCMagazine,2015- ieeexplore.ieee.org

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

6

• Múltiples capasocultas• Aprendizajejerárquico

• Característicascadavezmáscomplejas• Muybuencomportamiento enmútiplesdominios:Vision,Audio,…

DeepArchitecture(Redes Neuronales conmuchas capas)

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

7

UnsupervisedDeepArchitecture: Autoencoder

An autoencoder neuralnetworkisanunsupervisedlearningalgorithmthatappliesbackpropagation,settingthetargetvaluestobeequaltotheinputs.

Theaimofanautoencoder istolearnarepresentation(encoding)forasetofdata,typicallyforthepurposeofdimensionalityreduction.

Andrew Ng

Unsupervised feature learning with a neural network

x4

x5

x6

+1

Layer1

Layer2

x1

x2

x3

x4

x5

x6

x1

x2

x3

+1

Layer3

a1

a2

a3

Andrew Ng

x4

x5

x6

+1

Layer1

Layer2

x1

x2

x3

+1

a1

a2

a3

New representation for input.

Unsupervised feature learning with a neural network

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

10

DeepArchitecture: Autoencodern BreastCancerWisconsin(Diagnostic)DataSet

Autoencoder (f(x)=x) (unasolacapainternade3neuronasy1000"epochs".WDBC(569instanciascon32atributosdeentrada)

Credito:D.Charte

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

11

http://arxiv.org/abs/1312.5602

DeepMind:Start up-2011Demis Hassabis,Shane Legg yMustafa Suleyman

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

12

JuegosArcade(Breakout)

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

13

DeepLearningRetos enla“pintura”

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

14

http://www.deepart.io/

DeepLearning: DeepART

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

15

Ejemplos delresultado deDeepART

http://www.deepart.io/

vanGoth

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

16

Ejemplos delresultado deDeepART

http://www.deepart.io/

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

17

ModelodeDLutilizadoydescripcióndelametodología

http://arxiv.org/abs/1508.06576

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

18

Supervised DeepLearning:Convolutional NeuralNetworks

http://parse.ele.tue.nl/cluster/2/CNNArchitecture.jpg

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

19

Supervised DeepLearning:Convolutional NeuralNetworks

Eachmoduleconsistsofaconvolutional layerandapoolinglayer.

Typicallytriestocompresslargedata(images)intoasmallersetofrobustfeatures,basedonlocalvariations.

Basicconvolutioncanstillcreatemanyfeatures.

CNNshavebeenfoundhighlyeffectiveandbeencommonlyusedincomputervisionandimagerecognition.

DeepLearning

Convolutional steps

https://github.com/rasbt/python-machine-learning-book/blob/master/faq/difference-deep-and-normal-learning/convolution.png

http://neuralnetworksanddeeplearning.com/chap6.html

DeepLearningDigitRecognizerandConvolutional NN

Convolutional neuralnetworks usethreebasicideas:localreceptivefields,sharedweights,andpooling.

Localreceptive fields:Tobemoreprecise,eachneuroninthefirsthiddenlayerwillbeconnectedtoasmallregionoftheinputneurons,say,forexample,a5×5region,correspondingto25inputpixels.So,foraparticularhiddenneuron,wemighthaveconnectionsthatlooklikethis:

http://neuralnetworksanddeeplearning.com/chap6.html

DeepLearningDigitRecognizerandConvolutional NN

Localreceptive fields:

24×24 neurons

28×28 input image

http://neuralnetworksanddeeplearning.com/chap6.html

DeepLearningDigitRecognizerandConvolutional NN

Shared weights andbiases:thesame weightsandbiasforeachofthe24×24hiddenneurons(sigmoide function)

Themapfromtheinputlayertothehiddenlayerafeaturemap.

http://neuralnetworksanddeeplearning.com/chap6.html

DeepLearningDigitRecognizerandConvolutional NN

Shared weights andbiases:

In the example shown, there are 3 feature maps.

If we have 20 feature maps that's a total of 20×26=520 parameters defining the convolutional layer. By comparison, suppose we had a fully connected first layer, with 784=28×28 input neurons, 30 hidden neurons, 23,550 parameters.

http://neuralnetworksanddeeplearning.com/chap6.html

DeepLearningDigitRecognizerandConvolutional NN

Pooling layers: thepoolinglayersdoissimplifytheinformationintheoutputfromtheconvolutional layer,onecommonprocedureforpoolingisknownasmax-pooling,inthe2x2regioninput.

DeepLearningDigitRecognizerandConvolutional NN

Pooling layers:

http://neuralnetworksanddeeplearning.com/chap6.html

DeepLearningDigitRecognizerandConvolutional NN

The20imagescorrespondto20differentfeaturemaps

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

28

Credits:https://www.datarobot.com/blog/a-primer-on-deep-learning/

Alaizquierda,losdígitosdeentradasinprocesar.Aladerecha,representacionesgráficasdelascaracterísticasaprendidas.Enesencia,laredaprendea"ver"líneasybucles.

DeepLearning:MNISTdata

Asnapshotofimagepre-processingforconvolutional neuralnetworks:casestudyofMNISTSiham Tabik,DanielPeralta,AndrésHerrera-Poyatos,FranciscoHerreraInternationalJournal ofComputational Intelligence Systems,Vol.10(2017)555–56899.72accuracy

DeepLearningDigitRecognizerandpreprocessing

CNNmodels:LeNet,Network3,DropConnet

Preprocessing andaugmentation

Ensembles

DeepLearningDigitRecognizerandpreprocessing

CNNmodels:LeNet [4]

[4]Yann LeCun,Leon Bottou,Yoshua Bengio,andPatrickHaffner.Gradient-basedlearningappliedtodocumentrecognition.Proc.IEEE,86(11):2278– 2324,1998.

DeepLearningDigitRecognizerandpreprocessing

CNNmodels:Network3[13]

[13]MichaelANielsen.Neuralnetworksanddeep learning.URL:http://neuralnetworksanddeeplearning.com,2015.

DeepLearningDigitRecognizerandpreprocessing

CNNmodels:DropConnet [14]

[14]LiWan,Matthew Zeiler,Sixin Zhang,Yann LCun,andRobFergus.Regularizationofneuralnetworksusingdropconnect.InProceedingsofthe30thInternationalConferenceonMachineLearning(ICML- 13),pages 1058–1066,2013.

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

Translation.Theimageistranslatedanumberofpixelstowardagivendirection.

Centering.Toeliminatewhitecolumns/rows,andtoresize by scaling.

Rotation.Theimageisrotatedtoagivenangleθ.

Elastic deformation.Image pixels areslightly movedinrandomdirections,keepingtheimage’stopology.

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

Ensembles

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

The28handwrittendigitsmisclassifiedbyensemble-5ofNetwork3Thedigitbetween()representsthecorrectclass.The13digitslabeledwithasterisksarealsomisclassifiedbyDropConnect

Deep Learning: MNIST data Asnapshotofimagepre-processingforconvolutional neuralnetworks:casestudy ofMNISTSiham Tabik,DanielPeralta,AndrésHerrera-Poyatos,FranciscoHerreraInternationalJournal ofComputational Intelligence Systems,Vol.10(2017)555–56899.72accuracy

The28handwrittendigitsmisclassifiedbyensemble-5ofDropConnetThedigitbetween()representsthecorrectclass.The13digitslabeledwithasterisksarealsomisclassifiedbyNetwork3

DeepLearningDigitRecognizerandpreprocessing

Preprocessing andaugmentation

The13handwrittendigitsmisclassifiedbyensemble-5ofDropConnetandNetwork3

DeepLearningMXNet

http://mxnet.io/get_started/index.html

DeepLearningMXNet

http://mxnet.io/how_to/finetune.html

DeepLearningLearningfromscratchvs fine-tuning(GoogLeNet yVGG-16)

VGG16http://www.robots.ox.ac.uk/~vgg/research/very_deep/

https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md

##Informationname:16-layermodelfromthearXiv paper:"VeryDeepConvolutional NetworksforLarge-ScaleImageRecognition"caffemodel:VGG_ILSVRC_16_layerscaffemodel_url: http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodellicense:see http://www.robots.ox.ac.uk/~vgg/research/very_deep/caffe_version:trainedusingacustomCaffe-basedframework

VGGisaconvolutional neuralnetworkmodelproposedbyK.Simonyan andA.ZissermanfromtheUniversityofOxfordinthepaper“VeryDeepConvolutional NetworksforLarge-ScaleImageRecognition”.Themodelachieves92.7%top-5testaccuracyinImageNet ,whichisadatasetofover14millionimagesbelongingto1000classes.

K.Simonyan,A.ZissermanVeryDeepConvolutional NetworksforLarge-ScaleImageRecognitionarXiv technicalreport,2014https://arxiv.org/pdf/1409.1556.pdf

DeepLearningLearningfromscratchvs fine-tuning(GoogLeNet yVGG-16)

VGG16

Source:https://www.cs.toronto.edu/~frossard/post/vgg16/

Modelandpre-trainedparametersforVGG16inTensorFlowhttps://www.cs.toronto.edu/~frossard/post/vgg16/

Keras:https://keras.io/https://keras.io/applications/DeepLearninglibraryforTheano andTensorFlowKeras isahigh-levelneuralnetworksAPI,writteninPythonandcapableofrunningontopofeither TensorFlow or Theano.Itwasdevelopedwithafocusonenablingfastexperimentation. Beingabletogofromideatoresultwiththeleastpossibledelayiskeytodoinggoodresearch.

Models for image classificationwith weights trained on ImageNet

DeepLearningLearningfromscratchvs fine-tuning(GoogLeNet yVGG-16)

VGG16

DeepLearningLearningfromscratchvs fine-tuning(GoogLeNet yVGG-16)

GoogLeNethttps://research.google.com/pubs/pub43022.html

Weproposeadeepconvolutional neuralnetworkarchitecturecodenamedInceptionthatachievesthenewstateoftheartforclassificationanddetectionintheImageNet Large-ScaleVisualRecognitionChallenge2014(ILSVRC2014).Themainhallmarkofthisarchitectureistheimprovedutilizationofthecomputingresourcesinsidethenetwork.Byacarefullycrafteddesign,weincreasedthedepthandwidthofthenetworkwhilekeepingthecomputationalbudgetconstant.Tooptimizequality,thearchitecturaldecisionswerebasedontheHebbian principleandtheintuitionofmulti-scaleprocessing.Oneparticularincarnationofthisarchitecture,GoogLeNet,a22layersdeepnetwork,wasusedtoassessitsqualityinthecontextofobjectdetectionandclassification.

Going Deeper with ConvolutionsChristianSzegedy, Wei Liu, Yangqing Jia, PierreSermanet, ScottReed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, AndrewRabinovicharXiv technicalreport,2014https://arxiv.org/pdf/1409.4842.pdf

DeepLearningLearningfromscratchvs fine-tuning(GoogLeNet yVGG-16)

GoogLeNethttp://www.robots.ox.ac.uk/~vgg/research/very_deep/

GooglLeNet inCaffehttps://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet

GooglLeNet inKerashttps://gist.github.com/joelouismarino/a2ede9ab3928f999575423b9887abd14http://joelouismarino.github.io/blog_posts/blog_googlenet_keras.html

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

52

Deep Learning: Detección de Armas en Video

R. Olmos, S. Tabik, F. Herrera (UGR) Video:Skyfall,201252

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

53

Deep Learning: Detección de Armas en Video

R. Olmos, S. Tabik, F. Herrera (UGR) Video:Skyfall,2012

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

54

Underwaterimagepatchesofcorals

Cellclassificationfromcervixsmears Fingerprintidentification/classificationEjemplosdeaplicaciones

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

55

DeepLearningLectura:RecentOverview

http://www.nature.com/nature/journal/v521/n7553/full/nature14539.html

Ciencia deDatosyDeepLearning:Neuronasartificialesparaaprender

56

LeCum,Bengio yHinton,2015:“Ultimately,majorprogressinartificialintelligencewillcomeaboutthroughsystemsthatcombinerepresentationlearningwithcomplexreasoning.Althoughdeeplearningandsimplereasoninghavebeenusedforspeechandhandwritingrecognitionforalongtime,newparadigmsareneededtoreplacerule-basedmanipulationofsymbolicexpressionsbyoperationsonlargevectors”

Losavancesenelconocimientodelcerebroyelrazonamientohumanopermitirándiseñarnuevosparadigmasderepresentaciónyrazonamientocomplejo.