Top Banner

Click here to load reader

Нейронные сети: практическое применение / Наталия Ефремова (NTechLab)

Apr 16, 2017

ReportDownload

Engineering

ontico

  • N-TechLab

  • DeepLearning :

    , , , , , , ..

  • :

    : (CNN) (RNN)

  • :

  • ?UngerleiderandMishkin (1982)GoodaleandMilner(1992)

    ?

  • PrimaryVisualCortex

    InferotemporalCortex

    PosteriorParietalCortex

    LGN

    Retina

    SC

    Pulv

    LGN-lateralgeniculatenucleusPulv- pulvinarSC- superiorcolliculusGoodaleandHumphrey (1998)

  • HubelandWiesel(1968)

    V1

  • Kobatake,Tanaka(1995)

  • :

    IT V2-V4 V1

  • ,

    ,2D3D,

  • f

    T

    E

    c

  • Fischler and Elschlager, 1973

  • FACE

  • X W Y

  • X*W=Y

  • 2*W=6

  • Error=2*W- 6

  • Error

    Time

    1 2 3 4 5 6 7

    Error=|2*w 6| =0,001

    W=2.9999?

  • !7,650000,600000002GPU

  • 2010 :

    PASCALVOC(20)

    CALTECH101(101 )

  • 20

    14,197,12222

    www.image-net.org

  • 20

    14,197,12222

    www.image-net.org

  • 20

    2016competitionwinner:

    269layerR-CNNnetwork

    GatedBi-directionalCNNforObjectDetectionX.Zeng etal.,TheChineseUniversityofHongKong,Sensetime Group

  • Input Simplefeatures

    Moreabstractfeatures

    Mappingfrom

    featuresOutput

    (Goodfellow, 2016)

    input to layer

    a

    a. convolution stage: affine transformb. detector stage: non-linearity (rectified linear)c. pooling

    b c

    next layerconvolutional layer

  • convolution pooling convolution pooling

  • convolution pooling convolution pooling

    feedforward

    M.Zieler,CVSS2015

  • convolution pooling convolution pooling

    H.Lee etal.,2011

  • ?

    Cadieu etal.,2014

    ITcortex

    V4

    Zeiler&Fergus2013

    Krizhevsky etal.2012

    HMO

    Acuracy

    chance

    Cadieu etal.,2014

  • ,CNN

    (objectdetection) (semanticsegmentation) (facerecognition) (humanbodypartsrecognition) (semanticboundarydetection) (saliencyestimation) (surfacenormalestimation)

  • (Kokkinos, 2016)

    CNN

    q low-q mid-q high-levelvisiontasks

  • (Kokkinos, 2016)

    CNN

  • /

  • ? ?

  • NTechLab

    q 95%- 10K q 99%- q

  • NTechLab

    q 95%- 10K q 99%- q

  • NTechLab

    q 0,51q q 20500q

    q 95%- 10K q 99%- q

  • NTechLab

    q 0,51q q 20500q

    q 95%- 10K q 99%- q

    q web-,AndroidiOSq 250100q MefaFace

    MegaFaceFindFace.ru

  • q RNN: q q :

  • Kahouetal.,2015

    input

    Hiddenlayer

    Output

  • q q Atariq AlphaGo

    AndrejKarpathyblog (2016)

    V.Mnih 2015

  • J.Leitner (2014)

  • Levineet.al(2016)

  • M.Xieetal.,2016

  • 3D-

    (Kamnitsas etal.2016)

  • 3D-

  • !

    [email protected]b.com

  • :

    www.theguardian.comwww.nvidia.com www.image-net.orgBlaise Agera y Arcas:How computers are learning to be creative (TED talk)

  • ?

Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.