Cox Automotive - Digital Discrimination: Cognitive Bias in Machine Learning

Tools/Communities:
AI Fairness 360 Toolkit: http://aif360.mybluemix.net/
Model Asset Exchange: http://ibm.biz/model-exchange
IBM's Public Call for Code Competition: https://callforcode.org/
Maureen's team: IBM's Center for Open Source Data and AI Technologies: http://ibm.biz/codait-projects

Talk Sources:

Podcasts/Tweets
https://leanin.org/podcast-episodes/siri-is-artificial-intelligence-biased
https://art19.com/shows/the-ezra-klein-show/episodes/663fd0b7-ee60-4e3e-b2cb-4fcb4040eef1
https://twitter.com/alexisohanian/status/1087973027055316994

Amazon
https://www.aclu.org/blog/privacy-technology/surveillance-technologies/amazons-face-recognition-falsely-matched-28
https://www.openmic.org/news/2019/1/16/halt-rekognition

Google
https://motherboard.vice.com/en_us/article/j5jmj8/google-artificial-intelligence-bias

COMPAS
https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
https://www.technologyreview.com/s/612775/algorithms-criminal-justice-ai/

Data for Black Lives
http://d4bl.org/about.html
2019 Conference Notes: https://docs.google.com/document/d/1E1mfgTp73QFRmNBunl8cIpyUmDos28rekidux0voTsg/edit?ts=5c39f92e

Gender Shades Project
http://news.mit.edu/2018/study-finds-gender-skin-type-bias-artificial-intelligence-systems-0212
https://www.youtube.com/watch?time_continue=1&v=TWWsW1w-BVo
https://www.ajlunited.org/fight

Other resources referenced in this talk:
https://www.nytimes.com/2018/02/12/business/computer-science-ethics-courses.html
https://www.vox.com/science-and-health/2017/4/17/15322378/how-artificial-intelligence-learns-how-to-be-racist
https://www.engadget.com/2019/01/24/pinterest-skin-tone-search-diversity/

Transcript

1. Cox Automotive
   Digital Discrimination:
   Cognitive Bias in Machine
   Learning
   May 30th, 2019
   

   View Slide

2. Digital Discrimination:
   Cognitive Bias in Machine
   Learning
   2
   May 8, 2019 / © 2019 IBM Corporation
   

   View Slide

3. “A cognitive bias is a systematic pattern
   of deviation from norm or rationality in
   judgment. Individuals create their own
   "subjective social reality" from their
   perception of the input.”
   - Wikipedia
   3
   

   View Slide

4. May 8, 2019 / © 2019 IBM Corporation 4
   

   View Slide

5. Examples of bias in machine
   learning.
   5
   May 8, 2019 / © 2019 IBM Corporation
   

   View Slide

6. Google’s Cloud
   Natural
   Language API
   6 Image Credit: #WOCinTech
   

   View Slide

7. October 2017 - Google Natural Language API
   https://cloud.google.com/natural-language/
   7
   Source: https://motherboard.vice.com/en_us/article/j5jmj8/google-artificial-intelligence-bias
   

   View Slide

8. October 2017 - Google Natural Language API
   https://cloud.google.com/natural-language/
   8
   Source: https://motherboard.vice.com/en_us/article/j5jmj8/google-artificial-intelligence-bias
   

   View Slide

9. October 2017 - Google Natural Language API
   https://cloud.google.com/natural-language/
   9
   Source: https://motherboard.vice.com/en_us/article/j5jmj8/google-artificial-intelligence-bias
   

   View Slide

10. “We will correct this
    specific case, and, more
    broadly, building more
    inclusive algorithms is
    crucial to bringing the
    benefits of machine
    learning to everyone.”
    10
    

    View Slide

11. NorthPointe’s
    COMPAS
    Algorithm
    11 Image Credit: #WOCinTech
    

    View Slide

12. Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    May 2016 - Northpointe’s COMPAS Algorithm
    http://www.equivant.com/solutions/inmate-
    classification
    12
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

13. May 2016 - Northpointe’s COMPAS Algorithm
    http://www.equivant.com/solutions/inmate-
    classification
    Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

14. May 2016 - Northpointe’s COMPAS Algorithm
    http://www.equivant.com/solutions/inmate-
    classification
    Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

15. May 2016 - Northpointe’s COMPAS Algorithm
    http://www.equivant.com/solutions/inmate-
    classification
    Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

16. Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    16
    Black Defendant’s Risk Scores
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

17. Source: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
    17
    White Defendant’s Risk Scores
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

18. BLACK VS. WHITE
    DEFENDANTS
    ○ Falsely labeled black defendants as likely
    of future crime at twice the rate as white
    defendants.
    ○ White defendants mislabeled as low risk
    more than black defendants
    ○ Pegged Black defendants 77% more likely
    to be at risk of committing future violent
    crime
    18
    

    View Slide

19. 19
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

20. Amazon
    Rekognition
    20 Image Credit: #WOCinTech
    

    View Slide

21. July 2018 - Amazon Rekognition
    https://aws.amazon.com/rekognition/
    21
    Source: https://www.aclu.org/blog/privacy-technology/surveillance-technologies/amazons-face-recognition-falsely-matched-28
    

    View Slide

22. April 2019 - Amazon Rekognition
    https://aws.amazon.com/rekognition/
    22
    Source: https://www.openmic.org/news/2019/4/4/a-win-for-shareholders-amazon
    

    View Slide

23. Joy Buolamwini,
    Algorithmic Justice League
    Gender Shades Project
    Released February 2018
    23
    

    View Slide

24. View Slide

25. “If we fail to make
    ethical and inclusive
    artificial intelligence
    we risk losing gains
    made in civil rights
    and gender equity
    under the guise of
    machine neutrality.”
    25
    - Joy Boulamwini
    @jovialjoy
    

    View Slide

26. Solutions?
    What can we do
    to combat bias
    in AI?
    26
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

27. 27
    https://www.vox.com/ezra-klein-show-podcast
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

28. “Coders are the
    most empowered
    laborers that have
    ever existed.”
    28
    - Anil Dash
    @anildash
    

    View Slide

29. EDUCATION IS
    KEY
    29 Image Credit: #WOCinTech
    

    View Slide

30. https://www.nytimes.com/2018/02/12/business/computer-science-
    ethics-courses.html
    

    View Slide

31. Questions
    posed to
    students
    in these
    courses...
    Is the
    technology
    fair?
    How do you
    make sure
    that the
    data is not
    biased?
    Should
    machines
    be judging
    humans?
    31
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

32. 32
    https://twitter.com/Neurosarda/status/1084198368526680064
    

    View Slide

33. FIX THE
    PIPELINE?
    33 Image Credit: #WOCinTech
    

    View Slide

34. “Cognitive bias in
    machine learning is
    human bias on
    steroids.”
    34
    - Rediet Abebe
    @red_abebe
    

    View Slide

35. January 2019 - New Search Feature on...
    https://www.pinterest.com
    Source:
    https://www.engadget.com/2019/01/24/pinterest-skin-tone-search-diversity/
    

    View Slide

36. “By combining the
    latest in machine
    learning and inclusive
    product development,
    we're able to directly
    respond to Pinner
    feedback and build a
    more useful product.”
    36
    - Candice Morgan
    @Candice_MMorgan
    

    View Slide

37. TOOLS TO
    COMBAT BIAS
    37 Image Credit: #WOCinTech
    

    View Slide

38. Tool #1:
    AI Fairness
    360 Toolkit
    Open Source Library
    38
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

39. http://aif360.mybluemix.net/
    

    View Slide

40. http://aif360.mybluemix.net/
    

    View Slide

41. TYPES OF METRICS
    ○ Individual vs. Group Fairness, or Both
    ○ Group Fairness: Data vs Model
    ○ Group Fairness: We’re All Equal vs What
    You See is What You Get
    ○ Group Fairness: Ratios vs Differences
    41
    

    View Slide

42. View Slide

43. View Slide

44. Machine Learning
    Pipeline
    In-
    Processing
    Pre-
    Processing
    Post-
    Processing
    44
    Modifying the
    training data.
    Modifying the
    learning
    algorithm.
    Modifying the
    predictions (or
    outcomes.)
    

    View Slide

45. View Slide

46. http://aif360.mybluemix.net/
    Demos
    

    View Slide

47. https://github.com/IBM/AIF360
    AI Fairness 360 Toolkit Public Repo
    47
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

48. http://aif360.mybluemix.net/community
    AI Fairness 360 Toolkit Slack
    48
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

49. Tool #2:
    Model Asset
    Exchange
    Open Source Pre-Trained
    Deep Learning Models
    49
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

50. Step 1: Find a model
    ...that does what you need
    ...that is free to use
    ...that is performant enough
    50
    

    View Slide

51. Step 2: Get the code
    Is there a good implementation available?
    ...that does what you need
    ...that is free to use
    ...that is performant enough
    51
    

    View Slide

52. Step 3: Verify the
    model
    ○ Does it do what you need?
    ○ Is it free to use (license)?
    ○ Is it performant enough?
    ○ Accuracy?
    52
    

    View Slide

53. Step 4: Train the model
    53
    

    View Slide

54. Step 4: Train the model
    54
    

    View Slide

55. Step 5: Deploy your
    model
    ○ Adjust inference code (or write from
    scratch)
    ○ Package inference code, model code, and
    pre-trained weights together
    ○ Deploy your package
    55
    

    View Slide

56. Step 6: Consume your
    model
    56
    

    View Slide

57. Model Asset
    Exchange
    The Model Asset Exchange (MAX) is a one
    stop shop for developers/data scientists to
    find and use free and open source deep
    learning models
    ibm.biz/model-exchange
    57
    

    View Slide

58. ○ Wide variety of domains (text, audio,
    images, etc)
    ○ Multiple deep learning frameworks
    ○ Vetted and tested code/IP
    ○ Build and deploy a model web service in
    seconds
    58
    Model Asset
    Exchange
    

    View Slide

59. ibm.biz/model-exchange
    59
    

    View Slide

60. http://ibm.biz/model-exchange
    http://ibm.biz/max-slack
    Model Asset eXchange (MAX)
    60
    May 8, 2019 / © 2019 IBM Corporation
    

    View Slide

61. UPDATE TO
    THE GENDER
    SHADES
    PROJECT
    61 Image Credit: #WOCinTech
    

    View Slide

62. 62
    http://www.aies-conference.com/wp-content/uploads/2019/01/AIES-19_paper_223.pdf
    

    View Slide

63. 63
    https://www.ajlunited.org/fight
    

    View Slide

64. 64
    Photo by rawpixel on Unsplash
    No matter what it is our
    responsibility to build
    systems that are fair.
    

    View Slide

65. 65
    https://callforcode.org/
    

    View Slide

66. THANKS!
    Any questions?
    @Mo_Mack
    66
    

    View Slide