Machines are Learning: Bringing Powerful Artificial Intelligence Tools to Developers

WAJUG, Liège, November 20th, 2017

Have you always wanted to add predictive capabilities or conversational interfaces to your application, but haven’t been able to find the time or the right technology to get started?
Everybody wants to build smart apps, but only a few are Data Scientists. This session will help you understand machine learning terminology & challenges.

Transcript

1. Machines are Learning
   Bringing Powerful Artificial Intelligence to All Developers
   Danilo Poccia
   AWS Technical Evangelist
   @danilop danilop@amazon.com
   danilop
   

   View Slide

2. Credit: Gerry Cranham/Fox Photos/Getty Images
   http://www.telegraph.co.uk/travel/destinations/europe/united-kingdom/england/london/galleries/The-history-of-the-Tube-in-pictures-150-years-of-London-Underground/1939-ticket-examin/
   

   View Slide

3. 1939 London Underground
   Credit: Gerry Cranham/Fox Photos/Getty Images
   http://www.telegraph.co.uk/travel/destinations/europe/united-kingdom/england/london/galleries/The-history-of-the-Tube-in-pictures-150-years-of-London-Underground/1939-ticket-examin/
   

   View Slide

4. Data Predictions
   

   View Slide

5. Data Model Predictions
   

   View Slide

6. Model
   

   View Slide

7. http://www.thehudsonvalley.com/articles/60-years-ago-today-local-technology-demonstrated-artificial-intelligence-for-the-first-time
   1959 Arthur Samuel
   

   View Slide

8. Machine Learning
   

   View Slide

9. Machine Learning
   Supervised
   Learning
   Inferring a model
   from labeled
   training data
   

   View Slide

10. Machine Learning
    Supervised
    Learning
    Unsupervised
    Learning
    Inferring a model
    from labeled
    training data
    Inferring a model
    to describe hidden
    structure from
    unlabeled data
    

    View Slide

11. Reinforcement
    Learning
    Perform a certain
    goal in a
    dynamic
    environment
    Machine Learning
    Supervised
    Learning
    Unsupervised
    Learning
    

    View Slide

12. Driving a vehicle
    Playing a game
    against an opponent
    

    View Slide

13. Clustering
    

    View Slide

14. Clustering
    

    View Slide

15. Tip: Try topic modeling with your own emails ;-)
    Topic Modeling
    Discovering abstract “topics”
    that occur in a collection of documents
    For example, looking for “infrequent” words
    that are used more often in a document
    

    View Slide

16. Regression “How many bikes will
    be rented tomorrow?”
    Happy, Sad, Angry,
    Confused, Disgusted,
    Surprised, Calm,
    Unknown
    Binary
    Classification
    Multi-Class
    Classification
    “Is this email spam?”
    “What is the
    sentiment of this
    tweet, or of this social
    media comment?”
    1, 0, 100K
    Yes / No
    True / False
    %
    

    View Slide

17. Training the Model
    Minimizing the Error
    of using the Model on the Labeled Data
    

    View Slide

18. Validation
    How well is this Model working on New Data?
    

    View Slide

19. Be Careful of Overfitting
    

    View Slide

20. Be Careful of Overfitting
    

    View Slide

21. Be Careful of Overfitting
    

    View Slide

22. Better Fitting
    

    View Slide

23. Better Fitting
    

    View Slide

24. Different Models ⇒ Different Predictions
    

    View Slide

25. Labeled Data
    

    View Slide

26. Labeled Data
    70%
    30%
    Training
    Validation
    

    View Slide

27. Neural
    Networks
    

    View Slide

28. 1943 Warren McCulloch, Walter Pitts
    Threshold
    Logic
    Units
    

    View Slide

29. 1962 Frank Rosenblatt
    Perceptron
    

    View Slide

30. ∑
    w1
    w2
    w3
    wn
    w0
    =
    output
    weights
    (parameters)
    activation
    function
    input
    

    View Slide

31. f(∑)
    w1
    w2
    w3
    wn
    w0
    =
    weights
    (parameters)
    activation
    function
    output
    input
    

    View Slide

32. f(∑)
    input output
    

    View Slide

33. 1969 Marvin Minsky, Seymour Papert
    Perceptrons:
    An Introduction
    to Computational Geometry
    A perceptron can only solve
    linearly separable functions
    (e.g. no XOR)
    

    View Slide

34. f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    input
    layer
    hidden
    layer
    output
    layer
    input output
    Multiple Layers
    Lots of Parameters
    Backpropagation
    

    View Slide

35. Microprocessor Transistor Counts 1971-2011
    Intel Xeon CPU
    28 cores
    NVIDIA V100 GPU
    5,120 CUDA Cores
    640 Tensor Cores
    https://en.wikipedia.org/wiki/Moore's_law
    

    View Slide

36. LeCun, Gradient-Based
    Learning Applied to Document
    Recognition,1998
    Hinton, A Fast Learning
    Algorithm for Deep Belief
    Nets, 2006
    Bengio, Learning Deep
    Architectures for AI, 2009
    Advances in Research 1998-2009
    

    View Slide

37. “Stacks of differentiable
    non-linear functions
    with lots of parameters
    solve nearly any predictive
    modeling problem”
    —Jeremy Howard, fast.ai
    

    View Slide

38. Image
    Processing
    

    View Slide

39. f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    f(∑)
    output
    How to give images in input
    to a Neural Network?
    Photo by David Iliff. License: CC-BY-SA 3.0
    https://commons.wikimedia.org/wiki/File:Colosseum_in_Rome,_Italy_-_April_2007.jpg
    

    View Slide

40. Convolution Matrix
    0 0 0
    0 1 0
    0 0 0
    Identity
    Photo by David Iliff. License: CC-BY-SA 3.0
    https://commons.wikimedia.org/wiki/File:Colosseum_in_Rome,_Italy_-_April_2007.jpg
    

    View Slide

41. Convolution Matrix
    1 0 -1
    2 0 -2
    1 0 -1
    Left Edges
    Photo by David Iliff. License: CC-BY-SA 3.0
    https://commons.wikimedia.org/wiki/File:Colosseum_in_Rome,_Italy_-_April_2007.jpg
    

    View Slide

42. Convolution Matrix
    -1 0 1
    -2 0 2
    -1 0 1
    Right Edges
    Photo by David Iliff. License: CC-BY-SA 3.0
    https://commons.wikimedia.org/wiki/File:Colosseum_in_Rome,_Italy_-_April_2007.jpg
    

    View Slide

43. Convolution Matrix
    1 2 1
    0 0 0
    -1 -2 -1
    Top Edges
    Photo by David Iliff. License: CC-BY-SA 3.0
    https://commons.wikimedia.org/wiki/File:Colosseum_in_Rome,_Italy_-_April_2007.jpg
    

    View Slide

44. Convolution Matrix
    -1 -2 -1
    0 0 0
    1 2 1
    Bottom Edges
    Photo by David Iliff. License: CC-BY-SA 3.0
    https://commons.wikimedia.org/wiki/File:Colosseum_in_Rome,_Italy_-_April_2007.jpg
    

    View Slide

45. Convolution Matrix
    0.6 -0.6 1.2
    -1.4 1.2 -1.6
    0.8 -1.4 1.6
    Random Values
    Photo by David Iliff. License: CC-BY-SA 3.0
    https://commons.wikimedia.org/wiki/File:Colosseum_in_Rome,_Italy_-_April_2007.jpg
    

    View Slide

46. Convolutional Neural Networks (CNNs)
    https://en.wikipedia.org/wiki/Convolutional_neural_network
    

    View Slide

47. ImageNet Classification Error Over Time
    0
    5
    10
    15
    20
    25
    30
    2010 2011 2012 2013 2014 2015 2016 2017
    Classification Error
    CNNs
    

    View Slide

48. 2012 ImageNet Classification with Deep Convolutional Neural Networks
    

    View Slide

49. SuperVision: 8 layers, 60M parameters
    0
    

    View Slide

50. 2013 Visualizing and Understanding Convolutional Networks
    

    View Slide

51. View Slide

52. View Slide

53. How Do Neural Networks Learn?
    ?
    More generic and can be reused
    as feature extractor for other visual tasks
    Specific
    to task
    Cat
    Dog
    0
    

    View Slide

54. The Challenge For Machine Learning: Scale
    Aggressive migration
    New data created on AWS
    PBs of existing data
    Data
    

    View Slide

55. The Challenge For Machine Learning: Scale
    Tons of GPUs
    Elastic capacity
    Pre-built images
    Aggressive migration
    New data created on AWS
    PBs of existing data
    Data Training
    

    View Slide

56. The Challenge For Machine Learning: Scale
    Tons of GPUs and CPUs
    Serverless
    At the edge, on IoT Devices
    Tons of GPUs
    Elastic capacity
    Pre-built images
    Aggressive migration
    New data created on AWS
    PBs of existing data
    Data Training Prediction
    

    View Slide

57. Natural Language Processing Experiment: Topic Modeling
    EC2 Spot Instances
    1.1 Million vCPUs
    

    View Slide

58. Fulfilment & logistics Search & discovery Existing products New products
    Thousands Of Amazon Engineers Focused On Machine Learning
    

    View Slide

59. View Slide

60. View Slide

61. Machine Learning On AWS Today
    

    View Slide

62. Artificial Intelligence In The Hands Of Every Developer
    S E R V I C E S
    P L A T F O R M S
    E N G I N E S
    I N F R A S T R U C T U R E
    GPU CPU IoT Mobile
    Apache MXNet Caffe 2 Theano PyTorch CNTK
    TensorFlow
    

    View Slide

63. Early Detection
    Of Diabetic
    Complications
    

    View Slide

64. FDA-approved
    Medical Imaging
    

    View Slide

65. Sports Analytics
    

    View Slide

66. Autonomous Driving Systems
    

    View Slide

67. Real Time, Per Pixel Object Segmentation
    

    View Slide

68. Centimeter-accurate Positioning
    

    View Slide

69. Computation Knowledge Engine
    

    View Slide

70. View Slide

71. S E R V I C E S
    P L A T F O R M S
    E N G I N E S
    I N F R A S T R U C T U R E
    Amazon ML Spark & EMR Kinesis Batch ECS
    GPU CPU IoT Mobile
    Apache MXNet Caffe 2 Theano PyTorch CNTK
    TensorFlow
    Artificial Intelligence In The Hands Of Every Developer
    

    View Slide

72. S E R V I C E S
    P L A T F O R M S
    Vision
    Amazon Rekognition
    E N G I N E S
    I N F R A S T R U C T U R E
    Amazon ML Spark & EMR Kinesis Batch ECS
    GPU CPU IoT Mobile
    Apache MXNet Caffe 2 Theano PyTorch CNTK
    TensorFlow
    Artificial Intelligence In The Hands Of Every Developer
    

    View Slide

73. Mona Lisa
    (Leonardo da Vinci)
    

    View Slide

74. View Slide

75. View Slide

76. View Slide

77. Mona Lisa
    (Prado's version)
    

    View Slide

78. Portrait of
    Maddalena Doni
    (Raphael)
    

    View Slide

79. Bynder allows you to easily create, find and use content
    for branding automation and marketing solutions.
    With our new AI capabilities,
    Bynder’s software… now allows
    users to save hours of admin
    labor when uploading and
    organizing their files, adding
    exponentially more value.
    Chris Hall
    CEO, Bynder
    ”
    “
    With Rekognition, Bynder revolutionizes marketing admin tasks with AI capabilities
    

    View Slide

80. S E R V I C E S
    P L A T F O R M S
    Speech
    Amazon Polly
    Vision
    Amazon Rekognition
    E N G I N E S
    I N F R A S T R U C T U R E
    Amazon ML Spark & EMR Kinesis Batch ECS
    GPU CPU IoT Mobile
    Apache MXNet Caffe 2 Theano PyTorch CNTK
    TensorFlow
    Artificial Intelligence In The Hands Of Every Developer
    

    View Slide

81. Generate Lifelike Speech With Amazon Polly
    25 languages
    “The temperature in
    Milanis 16 degrees
    Celsius”
    “The temperature
    in Milan is 16˚C”
    Amazon
    Polly
    52 voices
    

    View Slide

82. aws polly synthesize-speech
    --text "It was nice to live such a wonderful live show."
    --output-format mp3
    --voice-id Joanna
    --text-type text
    output.mp3
    

    View Slide

83. Duolingo voices its language learning service Using Polly
    Duolingo is a free language learning service where users
    help translate the web and rate translations.
    With Amazon Polly our users
    benefit from the most lifelike
    Text-to-Speech voices
    available on the market.
    Severin Hacker
    CTO, Duolingo
    ”
    “ • Spoken language crucial for
    language learning
    • Accurate pronunciation matters
    • Faster iteration thanks to TTS
    • As good as natural human speech
    

    View Slide

84. ”
    “
    Royal National Institute of Blind People creates and
    distributes accessible information in the form of
    synthesized content
    Amazon Polly delivers
    incredibly lifelike voices which
    captivate and engage our
    readers.
    John Worsfold
    Solutions Implementation Manager, RNIB
    • RNIB delivers largest library of
    audiobooks in the UK for nearly 2 million
    people with sight loss
    • Naturalness of generated speech is
    critical to captivate and engage readers
    • No restrictions on speech redistributions
    enables RNIB to create and distribute
    accessible information in a form of
    synthesized content
    RNIB provides the largest library in the UK for people with sight loss
    

    View Slide

85. S E R V I C E S
    P L A T F O R M S
    Chat
    Amazon Lex
    Speech
    Amazon Polly
    Vision
    Amazon Rekognition
    E N G I N E S
    I N F R A S T R U C T U R E
    Amazon ML Spark & EMR Kinesis Batch ECS
    GPU CPU IoT Mobile
    Apache MXNet Caffe 2 Theano PyTorch CNTK
    TensorFlow
    Artificial Intelligence In The Hands Of Every Developer
    

    View Slide

86. Amazon Lex
    Speech recognition and natural language understanding
    Automatic speech recognition
    Natural language understanding
    “What’s the weather
    forecast?”
    Weather
    forecast
    Amazon Lex
    

    View Slide

87. Amazon Lex
    Speech recognition and natural language understanding
    “It will be
    sunny
    and 16C”
    Automatic speech recognition
    Natural language understanding
    “What’s the weather
    forecast?”
    Weather
    forecast
    Amazon Lex
    

    View Slide

88. “It will be sunny
    and 16 degrees
    Celsius”
    Amazon Polly
    Amazon Lex
    “It will be
    sunny
    and 16C”
    Automatic speech recognition
    Natural language understanding
    “What’s the weather
    forecast?”
    Weather
    forecast
    Speech recognition and natural language understanding
    Amazon Lex
    

    View Slide

89. ”
    “
    Finding missing persons:
    ~100,000 active missing
    persons cases in the U.S.
    at any given time
    ~60% are adults,
    ~40% are children
    • Motorola Solutions applies Amazon
    Rekognition, Amazon Polly and Amazon
    Lex
    • Image analytics and facial recognition
    can continually monitor for missing
    persons
    • Tools that understand natural language
    can enable officers to keep eyes up and
    hands free
    Motorola Solutions is using AI to help finding missing persons
    Motorola Solutions keeps utility workers connected and
    visible to each other with real-time voice and data
    communication across the smart grid.
    

    View Slide

90. S E R V I C E S
    P L A T F O R M S
    Chat
    Amazon Lex
    Speech
    Amazon Polly
    Vision
    Amazon Rekognition
    E N G I N E S
    I N F R A S T R U C T U R E
    Amazon ML Spark & EMR Kinesis Batch ECS
    GPU CPU IoT Mobile
    Apache MXNet Caffe 2 Theano PyTorch CNTK
    TensorFlow
    

    View Slide

91. There’s Never Been A Better Time To Build Smart Apps
    

    View Slide

92. https://github.com/danilop/security-camera
    

    View Slide

93. Machines are Learning
    Bringing Powerful Artificial Intelligence to All Developers
    Danilo Poccia
    AWS Technical Evangelist
    @danilop danilop@amazon.com
    danilop
    

    View Slide