O'reilly AI Conference

Transcript

1. San Francisco, CA
   17 - 20 September 2017
   Bramantyo Adrian
   

   View Slide

2. What
   “It is an AI Event that focus on real-world
   implementations.”
   GDP Labs Confidential
   

   View Slide

3. PUT AI TO WORK
   

   View Slide

4. 1. Building Unbiased AI
   2. Accelerating AI
   3. Active Learning and Transfer Learning
   4. Object Detection
   GDP Labs Confidential
   

   View Slide

5. Introduction
   +
   Internet Company
   IS NOT
   GDP Labs Confidential
   

   View Slide

6. Introduction
   +
   AI Company
   IS NOT
   GDP Labs Confidential
   

   View Slide

7. 1. Building Unbiased AI
   2. Accelerating AI
   3. Active Learning and Transfer Learning
   4. Object Detection
   GDP Labs Confidential
   

   View Slide

8. Diversity Crisis in AI
   “We’ve already seeing society’s racial and gender
   biases being encoded into software that uses AI
   when built by such a homogeneous group.”
   Building Unbiased AI
   GDP Labs Confidential
   

   View Slide

9. Building Unbiased AI
   GDP Labs Confidential
   

   View Slide

10. Building Unbiased AI
    GDP Labs Confidential
    

    View Slide

11. Goals
    Building Unbiased AI
    GDP Labs Confidential
    

    View Slide

12. Building Unbiased AI
    GDP Labs Confidential
    

    View Slide

13. Ignoring
    - Failing to max a product to different groups
    - Failing to attract potential users
    Building Unbiased AI
    GDP Labs Confidential
    

    View Slide

14. 1. Building Unbiased AI
    2. Accelerating AI
    3. Active Learning and Transfer Learning
    4. Object Detection
    GDP Labs Confidential
    

    View Slide

15. Accelerating AI
    GDP Labs Confidential
    

    View Slide

16. Manufacturing Automotive Healthcare Financial
    Transforming Many
    Industries
    Accelerating AI
    GDP Labs Confidential
    

    View Slide

17. CPU GPU FPGA ASIC
    Computational Substrate
    Accelerating AI
    GDP Labs Confidential
    

    View Slide

18. why
    Deep
    Learning ?
    Accelerating AI
    GDP Labs Confidential
    

    View Slide

19. Latest Model Not Always Relevant Model
    Accelerating AI
    GDP Labs Confidential
    

    View Slide

20. SSD model designed and optimized for PASCALVOC/MSCOCO dataset
    Know Your Model Provenance
    Accelerating AI
    GDP Labs Confidential
    

    View Slide

21. Smart hospital will generate over
    3,000 GB PER DAY
    Self driving cars will generate over
    4,000 GB PER DAY … EACH
    The average internet user will generate
    ~1.5 GB OF TRAFFIC PER DAY
    A connected plane will generate over
    40,000 GB PER DAY
    A connected factory will generate over
    1,000,000 GB PER DAY
    Invest in Data
    Manual
    annotations
    Training
    data
    Accelerating AI
    GDP Labs Confidential
    

    View Slide

22. 1. Building Unbiased AI
    2. Accelerating AI
    3. Active Learning and Transfer Learning
    4. Object Detection
    GDP Labs Confidential
    

    View Slide

23. Active and Transfer Learning
    Active
    Attain good learning
    performance without
    demanding too many
    labeled data
    GDP Labs Confidential
    

    View Slide

24. Motivation
    Active and Transfer Learning
    GDP Labs Confidential
    

    View Slide

25. Human Active
    AI
    Classifier
    Output
    Human
    Annotation
    Confident
    N
    ot Confident
    Active Learning
    Active and Transfer Learning
    GDP Labs Confidential
    

    View Slide

26. Transfer
    It’s very
    expensive to
    train deep
    learning
    Active and Transfer Learning
    GDP Labs Confidential
    

    View Slide

27. Transfer
    as
    feature
    extractor
    Active and Transfer Learning
    GDP Labs Confidential
    

    View Slide

28. Transfer
    as
    fine tuning
    Active and Transfer Learning
    GDP Labs Confidential
    

    View Slide

29. 1. Building Unbiased AI
    2. Accelerating AI
    3. Active Learning and Transfer Learning
    4. Object Detection
    GDP Labs Confidential
    

    View Slide

30. Computer Vision
    Object Detection
    GDP Labs Confidential
    

    View Slide

31. Traditional
    Machine Learning to
    Deep
    Learning
    Object Detection
    Feature Engineering
    challenges:
    - Human Scaling
    - Computational
    Scaling
    GDP Labs Confidential
    

    View Slide

32. Object Detection
    Feature Learning
    challenges:
    - Need large datasets
    - Model become
    complex
    Traditional
    Machine Learning to
    Deep
    Learning
    GDP Labs Confidential
    

    View Slide

33. Object Detection
    CIFAR, ImageNet
    PascalVOC
    MSCOCO
    KITTI
    GDP Labs Confidential
    

    View Slide

34. Convolutional Neural Network
    Object Detection
    GDP Labs Confidential
    

    View Slide

35. Object
    Detection
    Object Detection
    GDP Labs Confidential
    

    View Slide

36. Object
    Detection
    History
    Object Detection
    Region Proposed Approach
    Direct Classification
    Refined Classification
    GDP Labs Confidential
    

    View Slide

37. Region
    Proposed
    Approach
    Object Detection
    - Run through image to detect
    regions
    - Process regions extensively:
    - Bounding boxes
    - Predict class
    GDP Labs Confidential
    

    View Slide

38. Direct
    Classification
    Object Detection
    - BB and class are predicted
    directly with a single network
    GDP Labs Confidential
    

    View Slide

39. Refined
    Classification
    Object Detection
    - Add another branch that
    predict pixel-wise classes
    GDP Labs Confidential
    

    View Slide

40. Training
    Vision
    Model
    Challenges
    Object Detection
    GDP Labs Confidential
    

    View Slide

41. Training
    Vision
    Model
    Challenges
    Object Detection
    Batch
    size
    epochs Top-1
    Accuracy
    Hardware Cost ($) Time
    256 90 73.0% 1 DGX station 129,000 21h
    8192 90 72.7% 1 DGX station 129,000 21h
    8192 90 72.7% 32 DGX station 4.1 million 1h
    32K 90 72.4% 512 KNLs 1.2 million 1h
    GDP Labs Confidential
    

    View Slide

42. Object Detection
    

    View Slide

43. THANK YOU
    

    View Slide