csce 636 neural networks (deep learning)deep learning and neural network input neural network output...

23
CSCE 636 Neural Networks (Deep Learning) Lecture 1: Introduction to Deep Learning Anxiao (Andrew) Jiang

Upload: others

Post on 26-Jan-2021

30 views

Category:

Documents


0 download

TRANSCRIPT

  • CSCE636NeuralNetworks(DeepLearning)

    Lecture1:IntroductiontoDeepLearning

    Anxiao (Andrew)Jiang

  • AI,MachineLearning,andDeepLearning

    Coming up: Prerequisites forthecourse

  • Prerequisitesforthecourse

    • FamiliarwiththePythonprogramminglanguage• Basicbackgroundinmachinelearning,linearalgebra,calculus.

    Coming up: coursewebsite

  • Coursewebsite

    • http://faculty.cse.tamu.edu/ajiang/636.html

    Coming up: textbook

  • Textbook:DeepLearningwithPython(required)

    Coming up: textbook

  • Textbook:DeepLearning(recommended)

    Coming up: textbook

  • Textbook:DeepLearningQuickReference(recommended)

    Coming up: textbook

  • Textbook:NeuralNetworksandDeepLearning(recommended)

    Coming up: textbook

  • Textbook:LearningfromData(recommended)

    Coming up: deep learning andneural network

  • DeepLearningandNeuralNetwork

    NeuralNetworkInput Output

    Whatneuralnetworkisdoing:computing(oftentransformationoffeatures/representations,andmakingafinaldecision).

    Coming up: example oftransformation

  • Exampleoftransformation

    Coming up: example oftransformation

  • Exampleoftransformation

    Coming up: example ofdeepneural network (DNN)

  • ExampleofDeepNeuralNetwork(DNN)

    Coming up: example ofDNN(continued)

  • ExampleofDNN(continued)

    Coming up: example ofDNN

  • ExampleofDNN

    Coming up:whataneural network does: learnafunction

  • Whataneuralnetworkdoes:learnafunction

    NeuralNetworkx

    valueoff(x)

    Theneuralnetworklearnsthefunctionf(x),eitherexactlyorapproximately.

    Coming up:whatisaneuron

  • Whatisaneuron

    Coming up:whatisaneural network (NN)

  • Whatisaneuralnetwork

    Coming up: how totrainaneural network

  • Howtotrainaneuralnetwork

    NeuralNetworkx

    valueoff(x)

    Theneuralnetworklearnsthefunctionf(x),eitherexactlyorapproximately.

    1.Usealotof(input,output)pairstotraintheneuralnetwork.2.Adjustweightstominimize thedifferencebetween f(x)andtheneuralnetwork’spredicted valuesoff(x)

    Coming up: applications ofdeep learning

  • ApplicationsofDeepLearning

    • Computervision(smartcamera,robot,self-drivingcars,etc.)• Naturallanguageprocessing(machinetranslation,chatbot,etc.)• Gameplaying(alphaGo,videogames,etc.)• Createartorproducts(painting,music,poem,fashion,etc.)• Datastorageandtransmission(datacompression,transmission,etc.)• Financeandeconomy(trading,recommendation,economysurvey,etc.)• Healthcare(readX-raypictures,diagnosis,drugdesign,etc.)• Physics,business,education,smarthomes,etc.(Moreandmoreapplicationseveryday.)

    Coming up:whydeep learning now?

  • Whydeeplearningnow?

    • Hardware:GPU,CUDA,parallelcomputing• Data:ImageNetandmanymore• Algorithms:activationfunctions,weight initializationschemes,optimizationschemesfortraining.

    Coming up: homework (without submission)

  • Homework(withoutsubmission)

    ReadChapter2BeforeNextClass

    Coming up: Videos ondeep learning

  • Videosondeeplearning

    • Howweteachcomputerstounderstandpictures,https://www.youtube.com/watch?v=40riCqvRoMs• ArtisticStyleTransferForVideos,https://www.youtube.com/watch?v=Uxax5EKg0zA• ChopinMusicGenerationwithRNN(RecurrentNeuralNetworks)andDeepLearning,https://www.youtube.com/watch?v=j60J1cGINX4