Machine Learning の新機能✨ / What's New in Machine Learning

WWDC19で発表されたMLの新機能まとめスライドです。
このスライドを見ればWWDCのML関連の話題がだいたい掴めます。

- お断り -
このスライドは一般公開されているWWDC19のスライドのみ引用しています。

iOS de KANPAI !【WWDC 2019 報告会】
15 min

Transcript

1. .BDIJOF-FBSOJOHͷ৽ػೳ✨
   

   View Slide

2. !LPPPPPUBLF
   ͡ΜΉ
   ͓அΓ
   ͜ͷεϥΠυ͸Ұൠެ։͞Ε͍ͯΔ88%$ͷεϥΠυͷΈҾ༻͍ͯ͠·͢
   

   View Slide

3. View Slide

4. ηογϣϯ
   

   View Slide

5. 7JEFP63-
   IUUQTEFWFMPQFSBQQMFDPNWJEFPTQMBZXXED
   7JEFP
   ৄ͘͠஌Γ͍ͨ࣌͸
   

   View Slide

6. Summary
   Create ML Domain API Core ML 3
   7JEFP
   

   View Slide

7. Create ML
   $SFBUF.-
   .-ϞσϧΛ4XJGUͰ࡞Δ͜ͱ͕Ͱ͖Δ
   

   View Slide

8. Create ML
   ࠓ
   

   View Slide

9. Sound Activity Tabular
   Image Text
   Create ML
   ৽
   

   View Slide

10. <9DPEF><0QFO%FWFMPQFS5PPM><$SFBUF.->
    Create ML
    ৽
    

    View Slide

11. Create ML
    ৽
    υϥοάυϩοϓͰ؆୯ʹ
    

    View Slide

12. Image Classifier Object Detector
    

    View Slide

13. %&.0
    

    View Slide

14. ෯޿͍.-ϞσϧΛ
    ؆୯ʹ࡞Δ͜ͱ͕Ͱ͖Δྫ
    Create ML
    

    View Slide

15. Image Classifier Object Detector
    7JEFP
    

    View Slide

16. αΠίϩͷ໨ΛσΟςΫτ
    Dice Detection Model
    5
    5
    4
    4
    6
    

    View Slide

17. Sound Classifier
    7JEFP
    

    View Slide

18. ԻΛฉ͖෼͚Δ
    

    View Slide

19. Text Classifier Word Tagger
    7JEFP
    

    View Slide

20. จষͷײ৘εύϜτϐοΫΛ෼ྨ
    Text Classification
    Topic Classification
    Spam/Not Spam
    Sentiment Analysis !
    "
    Lorem ipsum dolor sit amet, consectetur
    adipiscing elit, sed do eiusmod tempor
    incididunt ut labore et dolore magna aliqua.
    Label 1
    Label 2
    Label 3
    Ut enim ad minim veniam, quis nostrud
    exercitation ullamco laboris nisi ut aliquip ex
    ea commodo consequat.
    Excepteur sint occaecat cupidatat non
    proident, sunt in culpa qui officia deserunt
    mollit anim id est laborum
    

    View Slide

21. Tabular Classifier Tabular Regressor Recommender
    7JEFP
    

    View Slide

22. Hardtrail
    Mountain
    Aspen Valley
    Pleasant Cove
    For You:
    Rocky Peak Cactus Ridge
    ͜ͷࢁΛߴධՁʹͨ͠ਓʹ
    ࣍ʹొΔࢁΛϨίϝϯυ
    

    View Slide

23. Activity Classifier
    7JEFP
    

    View Slide

24. "QQMF8BUDIͷηϯαΛར༻
    Device Sensors
    g
    a
    r
    Accelerometer Gyroscope
    Magnetometer
    Altimeter
    

    View Slide

25. ͲͷߦಈΛ͍ͯ͠Δ͔Λ෼ྨ
    Jogging
    Standing
    Gestures
    Gaming
    Golf
    Swimming
    

    View Slide

26. .-Y"3
    +
    7JEFP
    

    View Slide

27. "QQMF1FODJMͰखॻ͖จࣈࣝผ
    

    View Slide

28. "3͢͝Ζ͘ͱ૊Έ߹Θͤ
    

    View Slide

29. ✨؆୯ʹ࣮ݱͰ͖Δ✨
    

    View Slide

30. ͡ΜΉͷਪ͠
    ϑϦεϏʔͷ౤͛ํ෼ྨͷ%&.0
    Frisbee Motion Classifier
    Hammer Bowler
    Chicken Wing Backhand
    Forehand
    7JEFP
    

    View Slide

31. Summary
    Create ML Domain API Core ML 3
    

    View Slide

32. %PNBJO"1*
    ֶशࡁΈͷϞσϧѻ͑Δ
    

    View Slide

33. Face capture quality
    Sound analysis
    Speech Saliency
    Speech on Mac
    Sentiment classification
    Animal Detection
    Text Recognition
    Object Tracking
    Attention Saliency
    Document Camera
    On device speech
    Word Tagging
    Image similarity
    Face Landmark
    Image Similarity
    Landmark Detection
    Image saliency
    Rectangle Detection
    Text catalog
    Word Embeddings
    Image Classification
    NL Transfer learning
    

    View Slide

34. 7JTJPO'SBNFXPSL
    *NBHF
    7JEFP
    

    View Slide

35. ਓ͕Ͳ͜ʹண໨͢Δ͔Λֶश
    Attention and Objectness Based Saliency
    

    View Slide

36. લܠͱޙܠΛֶश
    Attention and Objectness Based Saliency
    

    View Slide

37. ˠը૾͔ΒॏཁͳΦϒδΣΫτΛநग़
    Bounding Boxes
    (0,0)
    

    View Slide

38. ը૾ͷྨࣅ౓Λܭࢉ
    Precision and Recall
    Recall
    Percentage of Target Class retrieved from entire library
    

    View Slide

39. 'BDF-BOENBSLTͷվળ
    76pt
    Face Landmarks
    65pt
    Single Confidence Score Confidence Score per point
    New
    

    View Slide

40. ਓؒݕ஌
    New Detectors - Human Detector
    

    View Slide

41. 
    New Detectors - Cat and Dog Detectors
    Cat
    Dog Dog
    Dog
    Cat
    

    View Slide

42. τϥοΩϯάਫ਼౓ͷվળ
    New Object Tracker
    

    View Slide

43. 7JTJPO'SBNFXPSL
    5FYU
    7JEFP
    

    View Slide

44. จࣈ͓͜͠
    

    View Slide

45. ଎͞WTਖ਼֬͞
    Fast Versus Accurate
    Fast Accurate
    Processing time Optimized for real-time Asynchronous processing
    Memory footprint Smallest Larger
    Support for rotated text Limited Broad
    Support for variety of fonts Limited Diverse font styles
    Accuracy for natural language Good Best
    

    View Slide

46. ద੾ͳύϥϝʔλઃఆΛ͢΂͠
    Use case
    • Read codes/serial numbers just like a barcode reader
    • Constrained camera usage
    • Interactivity is key
    request = VNRecognizeTextRequest(completionHandler: recognizeTextHandler)
    request.recognitionLevel = .fast
    

    View Slide

47. ӳޠͷΈ
    

    View Slide

48. /BUVSBM-BOHVBHF
    7JEFP
    

    View Slide

49. จষͷײ৘ղੳ
    Natural Language
    Sentiment Analysis
    I was so excited for
    the season finale,
    $
    Natural Language
    Sentiment Analysis
    but it was a bit
    disappointing.
    I was so excited for
    the season finale,
    $
    ☹
    

    View Slide

50. Sentiment Analysis
    Text Classification
    Sentiment Analysis
    -1.0 1.0
    Natural Language
    0
    “We had a not so fun time in Hawaii cause mom twisted her ankle.”
    -0.8
    Natural Language Text
    ෛPSਖ਼ͷײ৘ͷ਺஋Խ
    

    View Slide

51. 4QFFDIBOE4PVOE
    7JEFP
    

    View Slide

52. ΦϯσόΠεͰԻ੠ೝࣝ
    On-Device Device Support
    iPhone 6s and later iPad (5th generation) and later All
    

    View Slide

53. ೔ຊޠͳ͍
    On-Device Language Support
    English

    United States, Canada, 

    Great Britain, India
    Spanish

    United States, Mexico, 

    Spain
    Italian
    Brazilian
    Portuguese
    Russian Turkish Chinese

    Mandarin and Cantonese
    

    View Slide

54. Ի੠ೝࣝ
    ˙ൃ࿩଎౓
    ˙Ұ࣌ఀࢭ
    ˙੠ͷಛ௃
    ৽
    

    View Slide

55. ੠ͷಛ௃
    ˙+JUUFS໎͍ͷ͋Δ੠ʁ
    ˙4IJNNFSԒͷ͋Δ੠ʁ
    // Printing new results when recognizing pre-recorded audio
    if result.isFinal {
    let formattedString = result.bestTranscription.formattedString
    let speakingRate = result.bestTranscription.speakingRate
    let averagePauseDuration = result.bestTranscription.averagePauseDuration
    for segment in recognitionResult.bestTranscription.segments {
    let jitter = segment.voiceAnalytics?.jitter.acousticFeatureValuePerFrame
    let shimmer = segment.voiceAnalytics?.shimmer.acousticFeatureValuePerFrame
    let pitch = segment.voiceAnalytics?.pitch.acousticFeatureValuePerFrame
    let voicing = segment.voiceAnalytics?.voicing.acousticFeatureValuePerFrame
    }
    }
    ৽
    

    View Slide

56. Summary
    Create ML Domain API Core ML 3
    

    View Slide

57. $PSF.-
    7JEFP
    

    View Slide

58. Core ML 3
    Model
    Flexibility
    Model

    Personalization
    NEW
    

    View Slide

59. Neural Network layers
    100+
    αϙʔτ
    

    View Slide

60. ഑෍ϞσϧϚγϚγ
    

    View Slide

61. ΦϯσόΠεΞοϓσʔτ
    Update
    Training examples
    UPDATED
    NEW
    

    View Slide

62. User Differences
    Ñ
    ൚༻తͳϞσϧͰ͸ਏ͍࣌
    

    View Slide

63. Update Task
    Ö
    ❤
    á
    StickerClassifier
    Update
    Task
    CustomStickerClassifer
    UPDATED
    طଘͷϞσϧΛΞοϓσʔτ
    

    View Slide

64. On-Device
    Privacy Available
    No server
    ΦϯσόΠεͷྑ͞
    

    View Slide

65. 5VSJ$SFBUF
    1ZUIPOͰ
    $PSF.-ϞσϧΛ࡞Δ͜ͱ͕Ͱ͖Δ
    ϥΠϒϥϦ
    7JEFP
    

    View Slide

66. ୔ࢁͷը૾͕ඞཁ
    Original Approach
    ࠓ
    

    View Slide

67. Ұຕͷը૾Ͱ෺ମݕ஌
    Synthetic Data Augmentation
    ৽
    

    View Slide

68. Synthetic Data Augmentation
    ৽ Ұຕͷը૾Ͱ෺ମݕ஌
    

    View Slide

69. View Slide

70. 88%$ʹࢀՃ͢Δ͔೰ΜͰ͍Δਓ޲͚
    IUUQTOPUFNVLPPPPPUBLFOOFBF
    

    View Slide

71. !LPPPPPUBLF
    ͡ΜΉ
    

    View Slide

72. %F/"͸ɺຊΠϕϯτͷ಺༰ɺฒͼʹ͓٬༷͕ຊΠϕϯτΛ௨ͯ͡ೖखͨ͠৘ใ౳ʹ͍ͭͯɺ
    ͦͷ׬શੑɺਖ਼֬ੑɺ࣮֬ੑɺ༗༻ੑ౳ʹ͖ͭɺ͍͔ͳΔ੹೚΋ෛΘͳ͍΋ͷͱ͠·͢ɻ
    

    View Slide